reviews - ece.iisc.ac.inanurag/papers/anurag/kumar-etal10wsn-intrud… · indian institute of...

Indian Institute of Science,

Bangalore, India

Mote: A smart wirelessdevice, comprisingminiature sensors, a lowpower microprocessor, asimple digital radio trasceiver,and a small battery, all in acompact package.

REVIEWS

Wireless sensornetworks forhumanintruderdetection

The SmartDetect Project Teama

Abstract | In this paper we report on the outcomes of a research and demonstration project

on human intrusion detection in a large secure space using an ad hoc wireless sensor

network. This project has been a unique experience in collaborative research, involving ten

investigators (with expertise in areas such as sensors, circuits, computer systems,

communication and networking, signal processing and security) to execute a large funded

project that spanned three to four years. In this paper we report on the specific engineering

solution that was developed: the various architectural choices and the associated specific

designs. In addition to developing a demonstrable system, the various problems that arose

have given rise to a large amount of basic research in areas such as geographical packet

routing, distributed statistical detection, sensors and associated circuits, a low power

adaptive micro-radio, and power optimising embedded systems software. We provide an

overview of the research results obtained.aSmartDetect is a research and demonstration project funded by the ER&IPR Division of the Defence

Research and Development Organisation (DRDO), Government of India, during the period 2006-2010. Thefollowing project team contributed to the work presented in this paper: Anurag Kumar (Principal Investigator),P. Vijay Kumar (co-Principal Investigator); (the following are in alphabetical order within category) FacultyInvestigators: Bharadwaj Amrutur, G.K. Ananthasuresh, Navakanta Bhat, R.C. Hansdah, Malati Hegde, JoyKuri, Vinod Sharma, Y.N. Srikant, Rajesh Sundaresan; Project Staff: Tarun Agarwal, S.V.R. Anand, Pallav Bose,Vijay Dewangan, Shalini Keshavamurthy, A.V. Krishna, Pavan Kumar, Sharath Kumar, D. Manjunath, SundeepPatil, Poornima V.L., K. Aditya Prasad, Santosh Ramachandran, Anurag Ranjan, Subathra Sampath, JeenaSebastian, Vishwas Vasuki; Students: Pranav Agrawal, Sahebrao Sidram Baiger, Vinod Kumar Chouhan, TaposhBanerjee, Satyam Dwivedi, Abhishek Gupta, Santosh Hedge, Neeraj Kumar, Prachee Jindal, Premkumar Karumbu,Sambuddha Khan, Girish Krishnan, Syam Krishnan, Chaitanya U. Kshirasagar, K.P. Naveen, Mohan Rathod,Deepak Ravi, U. Raviteja, R. Abu Sajana, Ramanathan Subramanian, Thejas, Amulya Ratna Swain, LalithaVadlamani, Leena Zacharias.

1. IntroductionThis paper provides an overview of amultidisciplinary, multifaculty project carriedout in the Indian Institute of Science in the areaof wireless sensor networks for the detection ofhuman intruders into secure regions, such as thegrounds of high security buildings, a large industrialinstallation, or an airport.

A smart wireless sensor device comprisesminiature sensors, a low power microprocessor, a

simple digital radio transceiver, and a small battery,all in a compact package. It is expected that suchdevices (commonly referred to as motes after theprototypical device developed at the Universityof California, Berkeley) will become so powerefficient, that in conjunction with energy harvestingtechnologies, and energy efficient algorithms, theirbatteries could last for years. Figure 1 shows theTelosB mote (developed and marketed by CrossbowTechnology [75]) which was the device on which

Journal of the Indian Institute of Science VOL 90:3 Jul–Sep 2010 journal.library.iisc.ernet.in 347

SmartDetect: The name ofthe wireless sensor networksystem that we developedfor intrusion detectioninside a secured area.

REVIEW The SmartDetect Project Team

Figure 1: The Crossbow TelosB “mote” which was utilised fordeveloping the SmartDetect WSN.

the SmartDetect system described in this paper wasdeveloped.

In a typical application, such devices wouldbe deployed in the 100s or 1000s, in a plannedor random fashion (being strewn onto a largetract of land from an airplane, or embedded intobuilding structures as they are constructed). Onbeing initialised, the devices would discover eachother, and then self-organise into a multihopwireless (packet) network. They would then begin tosense the environment, and use various distributedalgorithms to schedule packet transmissionsbetween themselves, and carry out variouscommunication and local computation tasks, e.g.,to detect and identify an event (such as a fire oran intruder) and communicate this to an operatorat the edge of the network. Thus, these systemscan be viewed as easily deployable, self-configuring,embedded distributed smart instrumentation.

These systems can be seen to depend on thefollowing key technical elements:

1. Efficient and low cost microcontrollers, sensingand energy harvesting devices;

2. Support software (i.e., operating systems andcompilers) that is simple (yet provides therequired primitives), compact, energy efficientand power aware;

3. Distributed and energy efficient algorithmsfor self-organisation, scheduling ofpacket transmissions, locationing, time-synchronisation, sensor data processing (suchas quantisation, data compression, detection,estimation, identification, classification, andtracking), and system security.

The specific application addressed in this projectwas one of securing from human intrusion theperiphery and grounds of a large building, industrial

installation, or an airport. The perimeter of thegeographical area that needs to be secured could beseveral kilometers (e.g., a 1 km square area). Theactivities in such a situation would normally belimited to a few centrally located buildings, car parks,driveways. However, large parts of the groundswould see little activity, and may only sporadicallybe patrolled. The problem, therefore, is to deploy awireless sensor network in order to detect quicklyand to locate any abnormal activity in the normallyinactive grounds areas. Such networks would clearlyneed to be long-lived. Another problem is thatof securing the grounds around a building for ashort time-period when some sensitive activity istaking place in the building (e.g., the visit of aninternational dignitary to a hotel or guest house).Such a sensor network could be deployed by a specialtask force entrusted with the security of the visitor,and the network would be removed and stored oncethe visitor leaves.

Among the major projects that also addressedsimilar objectives, two notable ones are “A Line inthe Sand” project (see [6] and [7]) and the VigilNetProject (see [31]). “A Line in the Sand” was thename given to a field experiment conducted bythe NEST team of Ohio State University under theDARPA-NEST program. The experiment involvedthe deployment of a 90 node wireless sensor networkwith 78 magnetic sensors, and 12 additional radarsensors. A major contribution of the work was thatit demonstrated the feasibility of discriminationbetween object classes using a network of binarysensors. The VigilNet project, executed at theUniversity of Virginia, used magnetic sensors todetect and track the position of moving vehicles.Node power management was projected as themajor strength of this effort. This was performed bywhat the authors call “sentry service component”that selected a subset of motes called sentriesto monitor events. The remaining motes wereallowed to remain in a low-power state until anevent occurred. Both these projects used Mica2motes [75] under the TinyOS operating system.For the operation of the various protocols usedin these projects, a tight time synchronizationand maintenance of neighbour information wereessential. Both these requirements are hard tomeet, given the harsh outdoor environment wherenode failures are common, and the large networkdiameter. Also, message security, which is extremelyimportant for this kind of application, was notconsidered. Our major contribution is a design forSmartDetect that takes into account the above issuesand makes the solution robust, reliable, scalableand secure. SmartDetect does not assume tight timesynchronization, does not maintain any neighbourinformation, and has security features built into it.

348 Journal of the Indian Institute of Science VOL 90:3 Jul–Sep 2010 journal.library.iisc.ernet.in

Sensing modality: The typeof signal from theenvironment that is sensed bythe device. Examples includeacoustic, magnetic, optical,thermal, etc.

Wireless sensor networks for human intruder detection REVIEW

The SmartDetect project has been a uniqueexperience in collaborative research. It hasbrought together the expertise of ten facultymembers of varied interests (sensors andmicromechanical systems, circuits, computersystems, communication, networking, signalprocessing, and security), in order to executea large research and demonstration project thatspanned three to four years. After the identificationof the team and selection of the application area, theproject evolved as follows: (i) Key technical problemswere identified; (ii) Pieces of the puzzle, related toeach faculty’s expertise, were taken, analysed, andsolutions were developed; (iii) Various solutionswere implemented, tested, and rejected or revised.The project groups met twice a month over a periodof three to four years; once each month to discussamong themselves, and once to present relativelycomplete work to the entire project team, in whichmeeting the DRDO collaborators from CAIR (theCentre for Artificial Intelligence and Robotics) alsoparticipated. A website that contained the minutesof all project meetings and a repository of referredto and generated literature, was maintained. Broadlyspeaking, the project has resulted in two mainoutcomes:

• A demonstrable wireless sensor network forhuman intrusion detection, built on thecommercially available TelosB motes (seeFigure 1), with the human sensing modalitybeing passive infrared (PIR).

• A large amount of basic research leading topublications in journals and conferences, andtraining (in full or in part) of 10 PhD students,6 MSc (Engg.) students, 8 ME students, andover two dozen project staff.

In this paper we provide an overview of thetechniques that went into the development ofthe SmartDetect system, and also an overview ofthe related basic research that was conducted inareas such as sensors, low power radios for motes,geographical routing, distributed detection, andpower optimising systems software.

The following is the section-wise outline of thepaper. In Section 2, we first provide a comparativeoverview of sensor technologies. For this project,we found passive infra-red (PIR) sensors to bethe most effective. The remainder of Section 2describes the sensor platform design, and thesignal processing techniques used to infer humanpresence from the PIR sensor signals. In Section 3we discuss issues, and our solutions, related to thewireless mesh network that connects the PIR sensorplatforms to the base station. Topics discussed

are network self-organisation and geographicalforwarding. The processing environment on themotes is severely limited, with a very simpleoperating system and limited memory. In Section 4we describe the architecture of SmartDetect software,and some of the implementation challenges we faced.Being cheap devices, motes also have inaccurateclocks that drift relative to each other. In order tohave even a crude common notion of time, timesynchronisation becomes necessary; our approachto addressing this issue is provided in Section 5.In applications such as intrusion detection, thewireless sensor network is faced with adversarieswho jeopardise correct behaviour. Security protocolsare therefore necessary to thwart such attempts.Section 6 discusses SmartDetect’s robustness tocertain security attacks.

Our developmental efforts towards theSmartDetect platform have not only resulted inalgorithms and design insights, but have also ledto a considerable amount of new research results.In Section 7 we discuss the research on sequentialevent detection in sensor networks. Multimodalsensor fusion is a possibility that one could explorein the future. To this end, and with footstepsdetection in mind, a MEMS accelerometer, and itsassociated capacitance measurement electronics,have been developed; this is reported in Section 8.Innovative techniques are needed to reduce theenergy consumption of mote components. InSection 9 we report the development of a noveladaptive radio, that automatically adjusts itspower consumption depending on the qualityof the received signal. Operating systems provideapplications with a convenient interface to thehardware and compilers convert user programs tomachine code in an efficient manner. Research onan energy efficient operating system, and energyoptimising compilers is presented in Section 10.

2. Detecting human intrusion2.1. Intrusion sensors: An overviewHuman intrusion can be detected using many sensormodalities [6]. Some of the relevant ones are listedand compared in Table 1 (adapted from [6]). Sixtypes of sensors are included in the table. All ofthese are passive in the sense that, unlike radar orultrasonic sensors, they do not emit a signal andsense how targets modify it. Passive sensors arepreferred in sensor networks where there is limitedenergy. Magnetic sensors assume that the intruder,such as an armed person, has magnetically sensitivematerial. Ferromagnetic material creates a specificmagnetic signature that can be detected using amagnetometer. Any metallic content worn by theintruder can be detected using electromagnetic


Seismometers andgeophones measure grounddisplacements. Velometersare devices that measurevelocities.


techniques. Seismic and acoustic sensors are basedon the vibrations caused by the intruder. Both comeunder the general category of vibration sensorswhich we describe next in some detail as a preambleto our prototyping effort highlighted in Section 8.

Vibration-based surveillance sensors can beclassified into two major groups, namely, acousticsensors and motion sensors. Acoustic sensorsmeasure the sound produced by the entity thatis to be detected or monitored. In the case ofvehicles, the main sources of sound are engineand power-train noise, track/tyre noise and exhaustnoise. Footsteps of humans and animals, flutteringof wings by birds, etc., also generate sound inaddition to the entity’s vocal sound. Sensorsthat measure sound are essentially microphonesand hydrophones. On the other hand, vibratorymotion sensors sense displacement, velocityand acceleration using seismometers/geophones,velometers and accelerometers, respectively. Thephysical construction of both classes of sensors isalmost the same: they contain a spring-restrainedmass which inevitably will have some damping.However, the frequency, range of operation andresolution of these sensors will be significantlydifferent. Their cost also varies depending ontheir level of sophistication. It is unlikely that onesensor would work for detecting/monitoring variedsound/vibration sources.

Additionally, in the case of heavy vehicles theremight be coupling between the acoustic noise andground vibrations. The acoustic waves travel atdifferent speeds and their amplitudes decrease atdifferent rates with distance or get absorbed atdifferent rates. This helps in distinguishing thetype of vehicle or other noise source. Thus, in asurveillance application both acoustic and vibrationsensors are needed. A good example of this canbe found in an extensive study called “BochumVerification project for Military Vehicle Detection”[3] that was conducted to identify vehicles indifferent environments. The vehicles included in thissurvey were cars, small and large trucks, armoredpersonnel carriers and battle tanks. This study usedtwo different types of microphones, one type ofaccelerometer and one type of geophone in multiplenumbers. While an experiment on a tarmac road onsandy soil needed 1 accelerometer, 6 geophones and2 microphones, an experiment on a concrete roadon weathered layer of old lava needed 27 geophonesand 4 microphones. With multiple sensors, andsignal processing, the direction of arrival can beidentified.

Optical and thermal sensors work on theprinciple of disturbance in the line of sight ofthe sensors. Humans, animals, and vehicles have

‘hot spots’ or specific thermal signatures thatdistinguish them from vegetation and buildings,and enable detection. Chemical sensors rely onparticular chemical species associated with theintruder. Humans do leave a chemical trail butdetecting it requires the sophistication of traineddogs and warrants an array of specialized sensorsthat can detect many chemical species [6].

The criteria for comparison shown in Table 1were chosen keeping in mind their use in a sensornetwork. The comparison is subjective and dependson specific characteristics of particular sensors.

It is important that a sensor has sufficiently longrange so that the density of the sensor motes canbe kept reasonably low. Magnetic, thermal, andchemical sensors however have limited range andhence are less favoured than the other types.

When a sensor is used in a network for intrusiondetection, it is not enough to give a signal inthe event of intrusion; it is necessary to processthat information in order to avoid false alarms.For example, in a vibration sensor, not only themagnitude of the ground vibration but also thespectral (i.e., frequency related) information isnecessary to discern a disturbance as an intrusion.Sophisticated processing of the signals is needed. Agood sensor is one that requires the least processing.

Packaging and mounting or deploying an arrayof sensors is also an important consideration.Seismic sensors require more care than othertypes because the stiffness of their mountingsignificantly affects their performance. Acousticsensors, being sensitive to external vibrations, alsoare not favourable in this regard.

Sensitivity to line-of-sight obstructions can beeither good or bad depending on the application.But usually, a sensor that can detect in spite of anobstruction in between the sensor and the intruder isfavoured. A good example is a magnetic sensor thatis not affected by vegetation in between. Vibrationsensors too are not affected by an obstruction in theline-of-sight. On the other hand, optical, thermal,and chemical sensors get affected to various degreesby stationary or moving obstructions.

It is useful to have a sensor indicate the directionin which the intrusion has occurred. Magneticand thermal sensors cannot usually sense thisdirectionality of the intrusion.

In view of the limited energy available to a sensormote in a network, it is beneficial to have a provisionfor energy harvesting. Only vibration sensors areamenable to this. The same sensor element mightbe used to harvest vibration energy in the case ofseismic and acoustic sensors.

The last criterion listed in Table 1 is veryimportant. It is preferred that a sensor performs


Passive InfraRed (PIR) sensor:A sensor that measures theinfrared light emanating fromobjects. Motion detectorsusually use PIR sensors.


Table 1: Qualitative comparison of sensors for human intrusion detection. Here ‘+’ denotes desirable and ‘–’ denotesundesirable. (Adapted from [6])

Sensor Is it Extent of Level of Is it Is it Is energy Is ittype suitable signal complexity sensitive sensitive harvesting affected

for medium processing of to the to the possible? by theto long required: packaging line of direction ambientrange and sight? of the conditions?detection? mounting: intrusion

event?

Magnetic No (-) High (-) Low (+) No (+) No (-) No (-) No (+)

Seismic Yes (+) Medium (±) High (-) No (+) Yes (+) Yes (+) Somewhat (±)

Acoustic Yes (+) Medium (±) High (-) No (+) Yes (+) Yes (+) Somewhat (±)

Optical Yes (+) High (-) Low (+) Yes (-) Yes (+) No (-) Yes (-)

Thermal No (-) High (–) Low (+) Yes (-) No (-) No (-) Yes (-)

Chemical No (-) High (-) Low (+) No (+) No (-) No (-) Somewhat (±)

consistently in all ambient conditions such as lowor high winds, in bright light or in darkness, insunlight or in rain, etc. Only magnetic sensors aregood in this regard as compared with others.

Based on the foregoing discussion, it is apparentthat a single sensor might not completely meetthe needs of a surveillance application. It shouldalso be noted that commercial sensors, even ifthey are readily available and are within budget,may still pose packaging problems and may needconsiderable customization.

It is pertinent to note that military andenvironmental surveillance based on vibrationsensing using an array of micromachinedaccelerometers or microphones has attracted theattention of several groups. For example, DraperLabs has developed an array of biologically inspiredarray of microphones for localizing the direction ofsound within two degrees [20,55]. Applied MEMSInc. [5], has an Unattended Ground Sensor (UGS)module that uses micromachined accelerometer.Their device demonstrated that human footstepscould be picked up for monitoring purposes.Hougen et al. [36] used a vibration sensor as partof a miniature robotic system. Furstenau et al. [24]developed a vibration/acoustic sensor system formonitoring passage of vehicles, their direction andtype in an airport environment.

In this project, we chose passive infrared (PIR)sensors for initial demonstration of the efficacyof the algorithms and their implementation. Inview of long-term viability, seismic sensors werechosen for indigenous development. This wasbecause they score well on most criteria set forth inTable 1. However, the sensitivity and bandwidthrequirements are quite stringent when they areused in a surveillance application to detect human

footsteps. Such sensors, even though availablecommercially, are either very expensive or donot meet the performance requirements. Hence,the development of high-resolution and high-bandwidth accelerometer work was undertakenas part of this project. This is described in Section 8.

2.2. Utilizing off-the-shelf PIR sensorsTaking performance, pricing and availability intoaccount, the particular PIR sensor chosen for ourproject was the analog Panasonic motion sensorAMN24111 [62]. This sensor has four sensingelements arranged so as to form a 2×2 array (seeFig. 2) that is enclosed by a multilens made up ofmultiple contiguous plano-convex lenses [41,45].

Each single pixel is capable of capturingtemporal variations in the temperature. Thepixels are wired in differential mode to avoidtriggering the sensor due to changes in the ambienttemperature. Thus, any common temperaturechange simultaneously perceived by two pixels withopposite polarities fails to set off the sensor. Infrared(IR) radiation incident on each plano-convex lens isfocused onto the sensing elements of the PIR sensor.Thus, an array of virtual beams is created by the

Figure 2: The sensor and the quad pixels.


Clutter: Unwanted signals, inthe sensing modality, emittedby objects not of interest tothe detection problem.

Support Vector Machine(SVM): A supervised learningtechnique that attempts toseparate data points of onelabel from data points ofanother label, usually using ahyperplane in the data space.

Haar Transform (HT): Atransform representation ofthe original data, analogousto the Fourier Transform. TheHT has good time-frequencyresolution.


Figure 3: Cross-sectional and top view of the beams.

multilens which forms the field of view of the sensor.We refer to these beams as the Virtual Pixel Array(VPA). The top-view and cross-section of the VPAat a distance of 5 metres are shown in Fig. 3.

The sensor produces an electrical signalproportional to the difference in the rate of variationof incident–radiation–intensity between the twodiagonals (see Figure 2). When an intruder movesat a uniform speed along a circular path around thesensor, the beams are cut at a uniform rate. Thus,the sensor’s output resembles a triangular wave(sinusoidal after filtering) in which the numberof half-cycles is same as the number of beams inthe horizontal plane and the frequency observedis proportional to the angular velocity of theintruder1. The angular field view of each sensor isapproximately 110◦ and thus in our experimentalset-up, 3 sensors were mounted oriented at 120◦

relative to each other on a single platform, so asto obtain an omni-directional sensing range (seeFig. 2). The data from the three sensors were fed tothe 3 ADC channels of the TelosB mote.

The challenge and prior workWhile the response at the output of a PIR sensor

to a moving intruder is somewhat predictable,the principal challenge in using PIR sensors inan outdoor setting, is one of detecting intrusionsreliably in the presence of wind-blown clutterwhile maintaining a low false-alarm rate. Themanufacturers of commercially available sensorsfor instance, recommend careful placement of theirPIR detectors to prevent false alarms resultingfrom vegetation, air currents, etc. In [80,27,8],the PIR signal is first high-pass filtered to removelow frequency components resulting from slowenvironment changes and the signal energy isthen compared against an adaptive threshold. In

1These sensors were intended by Panasonic to be used asmotion detectors in indoor environments for applicationssuch as automated lighting etc.

Figure 4: Three Sensor Platform.

particular, an unsupervised adaptation techniqueis used to adjust the energy threshold for targetdetection. In most of the other work in the openliterature, decisions are made based on either simplethresholding of either the PIR signal itself or else,thresholding on the energy computed in a window.

Our approachWe have developed a low-complexity Support

Vector Machine (SVM) training-based algorithmthat uses the Haar Transform (HT) to separateintruder from clutter. We believe that training of theform inherent in SVM should form an essential partof any solution, given the large variations in intruderand clutter signatures. Also, SVM makes possiblea more fine-grain frequency-domain approach toseparating intruder from clutter. While results arepresented here only for the analog Panasonic MotionSensor AMN24111, they are readily extended toother PIR sensors. Training and testing were carriedout on actual experimental data and the algorithmwas found to exhibit good performance. The low-complexity nature of the algorithm serves to achievea key objective in a wireless sensor network, namelythe extension of network lifetime.

2.2.1. The detection algorithmWe assume the intruder to be a human traveling

in the vicinity of the sensor and use the term clutterto describe the sensor’s output as a result of wind-blown vegetation.



Figure 5: Spectral signatures of intruder and clutter at the output of the digital and analog PIRsensors respectively.

Figure 6: Functional block diagram of the algorithm.

Overview of the algorithmFor the purpose of maximizing battery life, we

decided to use the HT for computing the spectrumof intruder and clutter signals in preference tothe computationally intensive Discrete FourierTransform as only additions and subtractions sufficeto compute the HT. An alternative would have beento use Walsh Hadamard Transform (WHT) butthe HT was preferred as it can be computed withlower complexity, even when compared with theWHT. In addition, it has the ability to reuse pastcomputed HT coefficients for the next window andcan potentially be used to yield time-frequencylocalization information. A sampling frequency(fs) of 12.5 Hz was chosen based on the frequencycontent of intruder and clutter waveforms (seeFigure 5). A functional block diagram of thealgorithm appears in Figure 6. A block of 128 (N)

consecutive samples is transformed by the HT. Theenergy in each of these transformed componentsare binned into 8 frequency bins. The resultantbinned vector is passed on to a classifier (obtainedby off-line SVM training) which classifies it as eitherintruder or clutter. This entire process is repeatedon a sliding-window basis, every 16 (L) samples.

The Haar transform and frequency binningSince the Haar transform is wavelet based,

coefficients are designed to provide both frequencyand time localization information. As a result, thebreakdown of 128 Haar coefficients is as follows:there is one coefficient assigned to frequency 0 (theDC component) and 2k coefficients attached tosignals of frequency 2k , 0≤ k≤ 6. Thus, there are atotal of log(N )+1= 8 frequencies in all and in ouralgorithm, we collect together the energy in each of



Figure 7: Frequency bins corresponding to the 8-point Haar matrix.

these 8 frequency “bins”. The Haar signals associatedwith an example N=8-sample transform are shownin Fig. 7.

The time-localization aspect of the HT allowsreuse of components of a previously-transformedvector for the current window whenever thereis overlap of the current window with the priorwindow.

Support Vector MachinesThe SVM is a machine-learning technique

used for classification which when input with twolabeled sets of data (here binned vectors) returnsa decision surface which tends to maximize themargin between the two data sets [13]. Under linearSVM, the decision surface is chosen by SVM to bethe hyperplane in the input space that maximizesthe margin, i.e., the one that maximizes the distancebetween the hyperplane and the input data sets.In our case, the input space is 8 dimensional. Theoptimal hyperplane is typically found in practiceby reformulating it as a quadratic programmingoptimization problem which can be efficientlysolved. If the input data in linear SVM are not

Figure 8: Illustration of overfitting SVM with higher degree polynomialsin 2D.

separable by a hyperplane, we allow training errorsto occur. The tradeoff between the margin and thetraining errors can be controlled by a parameterC in SVM. The larger the value of C, the lesserthe number of training errors leading to a smallermargin. An alternative means of handling trainingerrors is to use a nonlinear separating surface. Inquadratic SVM, the optimal decision surface ischosen to be the hyperplane in a larger dimensionalspace which maximizes the margin from pointsin the larger dimensional space to which pointsin the data set are mapped. Each coordinate inthe larger dimensional space is associated to aunique monomial of degree 1 or 2 in the variablesattached to the input space. Thus the larger spaceis of dimension

(82

)+8+8= 44. One caveat is that

there is a risk of over-fitting the data with a highdegree separating surface that will perfectly separatethe training data but is likely to fail on any new data(see Fig. 8).

Computational complexitySince training of the SVM was done offline,

in estimating the computational complexity, weconsider only the computations carried out onlinein the mote on the incoming data. We considerlinear SVM classifiers initially and quadratic SVMclassifiers subsequently. SVM classification involvescalculating γ =wT x+b where w is the normal tothe hyperplane, b represents the affine shift of thehyperplane and x is the binned vector. Thus γ isproportional to the distance of x to the hyperplane.The number of computations required for the onlinepart of the algorithm for both linear and quadraticSVM appears in Fig. 9. Computations are in theorder of input size ‘N ’.

Training and testingThe data used for training SVM was collected

in a laboratory (i.e., clutter-free) environment (seeFig. 10) in the case of intruder and across 20 outdoorlocations on the forested campus of the IndianInstitute of Science (IISc) (see Fig. 11) in the case ofclutter. The training data set for intruder includesdata collected after making a human walk alongdifferent straight lines oriented in a variety of wayswith respect to the sensor. The clutter data wasaccumulated over the 6-month period from October2008 to March 2009. This would involve settingdown many nodes on the ground in selected clutter-prone areas and letting them observe and recorddata.

The data streams collected were partitionedinto two sets, one used for training, the other fortesting. The data streams from some of the morechallenging environments (for example, a sensor


False Alarm Rate: Theprobability that clutter ornoise are misclassified as anintrusion.

Miss Probability: Theprobability that an intrusionis misclassified as noise orclutter.


Figure 9: Computational complexity of the algorithm.

Figure 10: The indoor location used foraccumulating intruder data.

Figure 11: A location in IISc where a part ofclutter data was accumulated.

placed in the close proximity of large fern, or anintruder moving with high velocity) were usedfor training. From the data streams to be usedfor training, a total of 224 blocks were extracted,112 each representing intruder and clutter. Eachblock is a vector containing 128 consecutive timesamples. The offline training of SVM was done

using LIBSVM [17] in MATLAB. Linear SVM,with C set to a high value, was used. The trainingperformance recorded 7/112= 6.3% misses and4/112= 3.6% false alarms. It should be noted thatthe training set deliberately includes some datathat was hard to classify. Not surprisingly, testingperformance was significantly better than trainingperformance. The testing performance recorded3/500= 0.5% misses and 2000/160000= 1.25%false alarms. Some representative samples appearin Fig. 12. Fig. 12 shows on the left, four intrusionswhen the intruder sprinted four times in front ofthe sensor over a period of 70 s, all of which aresuccessfully detected by the algorithm. To the right,Fig. 12 shows clutter data collected over a period of80 s, the clutter was successfully rejected.

Limitations and future workAs mentioned, the data reported above was

for the period October 2008 to March 2009.However, when we carried out testing aroundnoon on a sunny day in April 2009, the heightof summer in Bangalore, we observed a significantlylarger false alarm rate. Fig. 13 shows a samplewaveform recorded in this period. When themid-summer noon’s data were also included inthe training set, linear SVM recorded a trainingperformance of 60/275 = 21.8% misses and22/275= 8% false alarms. Testing performancerecorded 30/300= 10% misses and 6100/100000=6.1% false alarms. Replacing the linear SVM with aquadratic SVM, improved the record on trainingdata to 47/275= 17% misses and 15/275= 5.5%false alarms. The improvement with regard to testingdata was far more pronounced. The use of multiplesensors may permit reduction of the false alarmrate since the false alarms were often caused by themovement of vegetation in close proximity to asensor. An alternative approach would be to detectintruders based on a better understanding of thesignature waveform of the intruder and clutter.

Field testingField testing was conducted on the lawns

of Electrical Communication Engineering (ECE)Department in the IISc campus. The initial decisionwas to deploy the sensor nodes in the form of a lineararray with inter-node spacing chosen to maximizethe area covered by a single node while ensuringthat every point in the sensing range was covered byat least 3 nodes. The idea here was that the sensingnodes would serve as a “wireless trip wire” (seeFig. 14). Larger areas can be covered by interlacingmany such wireless trip wires. It was found that asingle linear array would on occasion, fail to detectintruder moving at high-speeds, possibly because at


Scalability: The property ofan algorithm that enables itseasy adaptation tosituations with a largenumber of nodes.

ZigBee protocol: A simplifiedcommunication protocolbased on the IEEE 802.15.4standard for low data ratepersonal area networks.


Figure 12: Linear SVM: Intrusion detected on the left; clutter rejected to the right.

Figure 13: Quadratic SVM on summer clutter data.

Figure 14: Three sensors platform and thewireless trip-wire.

high speeds, the intruder was in the field of view ofa sensor for only a very short duration. We thereforedecided to create a double array comprising of two

identical, linear and parallel arrays spaced apart by5 m (the maximum sensing radius is 6 m). Decisionswere made locally as follows: If a node detected anintruder in its vicinity using the HT-cum-SVMbased algorithm outlined earlier, it would broadcastits local detection (via the ZigBee protocol availableon TelosB motes) to all of its neighbours. A nodewas permitted to declare a confirmed detection if, inaddition to making a local detection, it also receivednews of local detection from any other node withina distance of twice the sensing range of each sensor.The confirmed detection was then relayed back tothe base station using an appropriately designednetwork routing algorithm. At the base station, agraphical user interface (GUI) would display theinformation regarding the nodes that detected andthe route of the confirmed detection. This algorithmis scalable as the detection of an intruder resultsfrom the consensus of a few neighboring nodes.When tested over a period of several hours across theweek, the network performed flawlessly by detectingevery intrusion at speeds ranging from that of aslow crawl to a run at 5 m/sec. There were also nofalse alarms in the period over which testing wasconducted.

3. Alarm forwarding over a wireless meshnetwork

3.1. Node placementIn SmartDetect, the nodes can be functionallycategorized as sensor nodes and communicationnodes. Whereas the sensor nodes take part inevent detection, the communication nodes forwardinformation about detected events to the basestation. The sensor nodes are placed in two parallelrows so as to form a wireless trip-wire. The sensornodes are separated by a distance based on thesensing range of a PIR sensor, which is of theorder of a few meters. In our demonstration


Self-organisation: As appliedto communication networks,this is a means by which atopology or some sort oforder emerges in thenetwork. There is usuallylimited or no centralintervention, and all actionsare based on myopic ordistributed decision makingby participating agents.

Downstream nodes: The setof nodes that are closer tothe destination and reachablefrom a given node.


Figure 15: A typical deployment of sensor nodes.

setup communication nodes are placed in a semi-planned manner in which the area of deployment istessellated into square cells, and the nodes are placedwithin these cells randomly. This approach emulatesthe practical situation that due to the presence ofobstructions such as trees and taboo regions suchas ditches, it may not be possible to place nodesin a perfect grid. Work is currently under way todevelop methodologies for the design of wirelessmesh networks (which involves node placement, andtopology design) so that predictable performancecan be achieved in the delivery of information overwireless sensor networks (see [12]).

After the deployment, the base stationdisseminates location information to all the nodes.This information is later used for routing andlocalising the events. In the present implementation,the nodes are scheduled to stay awake all the time.Work on a sleep–wake cycling system is under wayat the time of writing this paper. The network canbe monitored from the base-station using a user-friendly graphical user interface (GUI). Figure 15shows a frame-grab from the GUI depicting 50nodes with communication nodes (filled circles),sensor nodes (filled squares on the top-left), and thebase station (BS).

3.2. Network self-organisationSelf-organisation is an essential step in the formationof an ad hoc wireless mesh network. In SmartDetect,self-organisation is done via a choice of transmitpower level at each node. Each node chooses theminimum power level so as to have a certain

minimum number of “good” neighbor nodes thatare closer to BS; these nodes will also be calleddownstream nodes. By “good” we mean that thesenodes can be reached with high packet receptionprobability. In our implementation, we require eachnode to have at least 2 downstream neighbours.

In [52], the authors provide a distributedalgorithm for constructing an approximateminimum spanning tree called a Nearest NeighborTree (NNT). The NNT algorithm bypasses the costlystep of cycle detection completely by a very simpleidea: each node chooses a unique rank, a quantityfrom a totally ordered set, and a node connects tothe nearest node of higher rank. This immediatelyprecludes cycles, and the only information thatneeds to be exchanged is the rank. The technique weused is an incremental neighborhood explorationsimilar to [52]. The Euclidean distance from theBS is used as the rank by the nodes in the self-organisation process. The nodes closer to the BShave smaller ranks than the ones farther away. TheBS has the least rank.

The self-organisation process works as follows.Each node in the network does this exploration oneafter the other in a time ordered fashion, startingfrom the base station. The base station starts theprocess by broadcasting request packets with theleast power. On the reception of a request packet, thereceiving nodes respond to the sender by sendingback an available packet using the same power levelthat is used by the sender of the request packet.The idea of sending the available packet using thesame power level used by the sender is to inform the


Geographic greedyforwarding: A routingprotocol in which packetsare forwarded to a locallyoptimal next-hop node withpositive spatial progresstowards the destination.


Figure 16: A self-organised sensor network with 50 nodes.

sender that they can communicate at that particularpower level. On reception of the available packet,the sender of the request packet stores the relevantinformation (for example, the number of availablepackets, the node ID, the power level, etc.).

After sending out a predefined number ofrequest packets, the node checks for goodneighbours in the stored information. If a sufficientnumber of such nodes is not discovered, the noderepeats this entire process with an increased power.This process ensures that by the end of the self-organisation process, every communication nodewill have multiple paths to the BS, while usingminimum power.

In order to identify downstream neighbours,each node compares its rank with that of itsneighbours. Figure 16 shows the result after self-organisation for a 50 node network. Nodes areshown in different colours to indicate the differentpower levels according to the table on the right side.After due consideration to the possibility of nodefailure and battery depletion, this self-organisationprocess can be scheduled at regular intervals tomaintain connectivity to the base station.

3.3. Alarm forwardingAlarm forwarding in the network involves routing ofthe alarm packets from the originating node to thesink. For such problems, geographic routing [39] is apopular protocol for packet delivery. Geographicrouting exploits the geographic informationinstead of topological connectivity informationto route packets to the destination. Geographic

routing protocols require only one-hop geographicinformation of neighboring nodes. The highlylocalized operation and the stateless feature ofgeographic routing make it simple and extremelyscalable. In geographic greedy routing, packets areforwarded to a locally optimal next-hop node witha positive progress towards the destination node.Such a protocol requires a node with a packet to beaware of its geographical location, and those of itsneighbours. This next hop node in the direction ofthe destination is called a relay or greedy node.

Battery-operated networks adopt radio duty-cycling [15,77], a MAC layer technique, to improvethe lifetime of the network. Here each nodeperiodically cycles between an awake state (radioON) and a sleep state (radio OFF). This resultsin time varying connectivity among nodes andaffects the maintenance of one-hop neighbours’geographic information required for greedy routing.Further, the radio duty-cycling leads to the optimalrelay node selection problem where there is atrade-off between the delay in relay node selectionand the progress made towards the sink. Thusit is imperative to develop a geographic greedyrouting protocol that provides a framework forincorporating the optimal relay selection algorithmwithout the maintenance of one-hop neighborinformation.

We propose Geography-aware MAC (GeoMac),a transmitter initiated geographic greedy routingprotocol. Here a node with a packet to forwardinitiates handshakes by broadcasting probes atregular intervals in order to determine greedy nodes.


Pseudo-events: Virtualevents (such as a fire-drill),created across space andtime, that monitor thehealth of the network.

TinyOS: A component-basedevent-driven open sourceoperating system for motes.


Figure 17: The failure of geographic greedy routing. S is a stuck node.

The probes contain the forwarding metric (distancesor hop counts) and the ID of the node broadcastingit. Nodes closer to the destination respond with aprobe ACK which contains its ID and forwardingmetric. An optimal timer starts on the reception ofthe first probe ACK. All the received probe ACKswithin the optimal time period are queued. Thesender now unicasts the packet to the relay nodeclosest to the destination. The value of the optimaltimer determines the relay node to which the packetwill be forwarded.

Geographic greedy forwarding may fail when allthe neighboring nodes of a sender are farther awayfrom the destination node than the sender itself.Figure 17 presents the local minimum condition(or a communication void [24]) where nodeS does not have any greedy relay nodes in itscommunication neighborhood. S fails to locatea next hop node in its neighborhood that has apositive geographic progress towards the destinationnode. Although a dense deployment of nodes canreduce the likelihood of occurrence of a void inthe network, packets may encounter voids withnonzero probability and cannot reach the sink. Thisis undesirable for WSNs deployed for applicationssuch as intrusion detection. Hence, there is a needto develop fault tolerant techniques that makes thenetwork resilient to communication voids.

Solutions such as face routing [39], convexhull routing [47] and link reversal routing [25]were proposed in the literature to pull thenetwork out of a local minimum condition. Allthese algorithms require knowledge of one-hopneighbours. Maintenance of one-hop neighborinformation in duty-cycled sensors involves severalmessage exchanges between nodes and its one-hopneighbours, associated access issues and collisionresolution mechanisms. This can be both time andenergy consuming.

In geographic greedy routing, packets are alwaysforwarded to nodes that make positive progresstowards the sink. Clearly, the resulting routing graph

is a directed acyclic graph (DAG). A DAG is saidto be destination-oriented when there is a directedpath in the DAG from any node to the sink [25]. ADAG is destination-disoriented if there exists a nodeother than the sink that has no outgoing links. Sucha node is said to be stuck. A destination-orientednetwork under geographic greedy routing may berendered destination-disoriented in the presenceof communication voids. Gafni and Bertsekas in[25] propose two general classes of link reversalalgorithms for solving the above problem in aneighbor aware fashion. The link reversals areachieved by distributed relabeling. We provided twoneighbour oblivious algorithms that stay within theframework of [25], and, thus have all the desirableproperties of the algorithms reported therein. Ourneighbour oblivious algorithms do not need to storeneighbour information, but learns it as and whenneeded.

In a sleep-wake cycling network, however,carrying out the void removal algorithm at thetime of an alarm incurs large alarm forwardingdelays [16]. We proposed the use of pseudo-eventsto maintain the network in a destination-orientedstate before the onset of real events. Pseudo-eventscreate virtual events distributed across space andtime. This initiates link reversal at stuck nodes whilerelaying these pseudo-alarm packets to the sink.Our proposed maintenance technique is likely tobe more energy efficient and works well even in aduty-cycled network since the forwarding protocolitself repairs the network.

4. SmartDetect implementation4.1. Software architectureSmartDetect runs on TinyOS-2.1 operating systemand the code is written in the nesC language. TinyOSis a component-based event-driven open sourceoperating system. The application code is built usingthe basic building blocks provided by the TinyOScore. The TinyOS core provides components formanaging timers, performing radio communication,collecting sensor samples from the analog-to-digitalconverter (ADC), scheduling various other tasksand so on.

Figure 18 depicts the complete softwarearchitecture of SmartDetect along with theTinyOS components it uses. The applicationcode is modularized and provides flexibility toreplace existing modules with others of the samefunctionality.

The application flow is as follows. The functionsof the network start with the initialization of nodesin the main application code. The base stationsends location information and other parametersnecessary for network operations to all the nodes.


Over-the-air programming:A method by which softwareupgrades are receivedwirelessly by the motes.


Figure 18: SmartDetect Software Architecture.

The nodes start the self-organisation process oncethey receive their coordinates. At each node, theoutput of this process is the neighbor list and apower level (just sufficient for communicationdirected towards the base station). This powerlevel will be set by the node and used for furthercommunication.

Once the self-organisation process completes,the sensors are initialized and they are sampledat regular intervals of 80 ms. The sample valuesobtained are fed to the detection module whichperforms the first level of processing by using theHaar-SVM algorithm (see Section 2.2.1). The resultof this algorithm is the decision of whether there wasan intruder or not. Based on this result, distributedalgorithms for event declaration are run in the mainapplication code.

Figure 19: TelosB program memoryutilisation in the SmartDetect implementation.

The packets that are to be sent to the base stationare handed-over to the Geo-MAC module whichforwards them to the base station. At each nodealong the way, the security module is responsiblefor encrypting and decrypting the messages whileforwarding. Only successfully decrypted messagesare passed on to the main and routing modulesfor further processing. Once the packets reach thebase station, they are handed over to the host over awireline serial communication interface.

The information received by the base stationregarding the event is displayed on a Java basedGUI. We have recently incorporated a networkmonitoring module which provides informationabout the health of the network. The informationobtained by querying the nodes is displayed on theGUI. For troubleshooting purposes, we have alsointroduced mechanisms for requesting and gettingstate information pertaining to a software modulewithin a node.

4.2. Implementation challengesThe program memory available on the TelosB moteis only 48 KBytes. This limited memory has posedrepeated challenges during the development ofSmartDetect. The problem is aggravated by thefact that half of this memory is used up by theTinyOS operating system, as shown in Figure 19. Inorder to accommodate all functionalities requiredby SmartDetect, we made careful design choiceswhile implementing the various algorithms, andcarefully optimized the code.

One of the interesting challenges we encounteredwas in the context of “over-the-air programming,”a very essential feature required to update thesoftware image running on the motes in-situ. Forthis, TinyOS has a built-in support known asDeluge. We found that incorporating the Delugecomponent required 40% of the program memory.We therefore came up with a scheme involving atwo step process. We make use of two applicationimages, an application image with Deluge alone,the other being our main SmartDetect application.We temporarily move from SmartDetect to theDeluge application during the software upgrade,and “reboot” back to SmartDetect after the upgrade.

5. Time synchronisationTwo clocks never agree all the time. Disagreementsarise because oscillators that drive the clocks inmotes are cheap, have varying characteristics fromdevice to device, and are sensitive to changesin the environment, particularly temperature.But synchronisation, even if only obtained toa certain degree, can indeed be useful. First,WSNs are severely energy constrained. Sleep-wake cycling of motes enhances their life time.


Skew and offset: Theparameters (slope andintercept) of a simple affinemodel, for relating the localtime at a node with that atthe reference node.

Probabilistic method of keydistribution: This refers to arandom assignment of acertain number of keys froma common pool.


Synchronised sleep-waking can allow a sufficientlysmall waking time just enough to compensatefor minor drifts, yet sufficient to communicatewith a neighbouring sensor during the awake time,thereby increasing life time of the sensor and thenetwork. Second, when tracking a moving object,time-stamped data from various sensors have tobe appropriately sequenced for reconstruction ofthe intruder’s path. Synchronisation is essential toachieve this sequencing. Third, secret keys for securecommunication are frequently changed to improvesecurity. Synchronisation ensures that these keychanges are coordinated.

Several algorithms for synchronisation areavailable. The most common method is to makea reference node broadcast a packet with its timeinformation. If propagation delays are negligible,this broadcast enables nodes to calibrate their timeline with this common reference tick, and makecorrections as needed. However, this only exploitsthe star-like connections from the reference node tothe other nodes. Links between other nodes are notexploited. Solis, Borkar, and Kumar [72] providea different algorithm that exploits these connections.In particular, the net clock offset around any loopis zero, analogous to the well-known Kirchchoff ’svoltage law. Similarly, the net product of the skews is1, yielding another conserved quantity. These ideasprovide a distributed and asynchronous algorithmfor clock updates. The algorithm also adapts tochanges in network as long as the network is able toelect one of the nodes as a leader (reference node),i.e., it is robust to a change in the leader. A roughdescription of the two-stage algorithm is as follows.All sensors track their skews and offsets with respectto a chosen reference node.

(1) In the first stage a node conducts bilateralexchanges of packets containing transmit andreceived time stamps with each of its neighbours.This enables it to estimate its relative skew andrelative offset with respect to each neighbour.

(2) In the second stage this node hears broadcastinformation of neighbours’ skews and offsets withrespect to the global reference node. An estimateof the node’s offset with respect to the reference isthe sum of a neighbour’s offset and the computedlink-level offset with this neighbour. Thus eachneighbour gives an estimate of the node’s offset. Theupdated offset is the average of all these estimates.

A similar calculation is done for skews withskews and offsets computed treating the otherquantity as known. The above algorithm is robustto distributed and asynchronous updates. It canalso handle changes in the reference node quitegracefully. The new reference node simply stopsmaking updates, and the old reference node begins

making updates. The converged values will nowindicate offsets and skews with respect to the newreference node.

The above algorithm is a discrete-version ofthe so-called averaging principle that underpinsseveral physical phenomena such as heat transferin a medium and social phenomena such as spreadof gossip in a network of individuals. Simulationsindicate that the algorithm can provide tens ofmicroseconds accuracy, with the limitation factorbeing the accuracy of the time-stamping mechanismitself.

6. Network securityGiven that the network may need to operate inremote and potentially hostile environments, securemessage exchange is a basic need.

Secure message exchange imposes severalrequirements, including

• Encryption: Transmissions are encrypted, sothat an adversary capturing packets off the airwould see only ciphertext, and not plaintext

• Authentication: A sensor node receiving amessage needs to be sure that the messagewas transmitted by a friendly neighbour, andnot a malicious adversary

• Message Integrity: Similarly, one would liketo ensure that a received message has notbeen tampered with, and its contents havenot been modified in any way.

Public-key or asymmetric cryptography andprivate-key or symmetric cryptography are the twochoices available for implementing a secure system.Public-key cryptography imposes significantlyhigher computational requirements. Keeping inmind the limitations of the off-the-shelf sensor nodeplatforms to be used in the prototype system, theoption of private-key cryptography was selected.

6.1. New key distribution algorithmsKeys can be provided to sensor nodes beforedeployment, or they can be generated by nodes afterdeployment in the field. In the following paragraphs,we provide an overview of a few algorithms of eachtype that have been developed in the course of theproject.

When keys are provided before deployment,there are two strategies that can be followed. Inthe probabilistic method, a random selection ofk keys (say) is made from a pool of P keys, andassigned to a sensor. In this way, a random subsetis chosen independently for each sensor. Thus,after deployment in the field, two nodes wishing tocommunicate may or may not have keys in common.


Friends of a given node arethose nodes that share acommon key with the givennode.


If they share one or more keys, then they are able tocommunicate securely. If they do not share keys,then the nodes need to execute some algorithm toobtain shared keys.

In [29], we propose the “Friends”-basedalgorithm to generate common keys between asource S and a destination D. Essentially, S initiatesa search for nodes X that have common keys withD; these are the friends of S. Each X sends a part ofa common key back to S. S constructs a full key bya random choice from the part-keys received, andinforms D about the identities of those X whosepart-keys were used. This enables D to construct thesame full key that S has, and henceforth, the twocan communicate securely.

Evaluation of the proposed algorithm considersthe feasibility of finding enough friends for thescheme to work, as well as a measure for quantifyinghow secure the generated key is — the probabilitythat an adversary is able to recover the generatedkey in L attempts. A third metric used is the energyoverhead of the method [30]. Analytical expressionsare obtained for each metric, and comparisonswith a related scheme in the literature [50] yieldencouraging results.

The second method of providing keys beforedeployment is a deterministic method, in whichkeys are allocated to nodes explicitly. In [28], wepropose a method in which a pool of p2 keys (p:prime) is arranged in p×p grid, and a sensor nodecorresponds to a polynomial over the residue field{0,1, . . . , (p−1)}. The polynomial passes throughpoints in the grid, and these keys are allocated tothe node. The idea is a generalization of [38].

Evaluation of the algorithm considers theconnectivity and resilience metrics. Connectivityrefers to the probability that two randomly chosennodes have at least one common key between them.Resilience indicates the fraction of keyed links thatare compromised when c nodes (say) are capturedby the adversary. Analysis yields formulae for themetrics, and suggests design rules in terms of howto choose p in order to achieve desired values ofconnectivity and resilience. A particular case ofthe deterministic algorithm is the full connectivityscheme, in which connectivity is guaranteed. Ourwork in [28] includes an algorithm for assigningkeys such that full connectivity is ensured.

Next, we consider algorithms for key generationafter nodes have been deployed. The topic hasreceived a lot of attention in the literature, anda recent trend is to examine if knowing the expectedlocations of nodes helps in improving connectivityand security. In [2], we propose an algorithmfor generating keys in situ, after nodes have beendeployed. The algorithm is initiated by a few nodes

called “tagged nodes,” that are programmed totransmit a few broadcast packets after deployment.Essentially, each tagged node can be thoughtof as a leader with its own pool of keys, anddifferent tagged nodes have disjoint key pools.Normal nodes — which are exactly the same astagged nodes except that they do not send initialbroadcasts — receiving messages from leadersassociate themselves to a chosen leader’s group(some nodes choose more than one leader). Whena few nodes are compromised, the revealed keysaffect communication among relatively few nodes;this is because sensors associated with distincttagged nodes use distinct key pools. The idea isto restrict the impact of loss of keys to a limitedarea, rather than the complete network. Simulationsindicate that both connectivity and security improvecompared to schemes that leverage expected nodelocations; this is interesting, because the proposedscheme does not use knowledge of node locationsin any way.

6.2. Implementation on off-the-shelf devicesImplementing security on commercial off-the-shelf(COTS) sensor motes posed its own set of challenges.As indicated earlier, the TinyOS operating systemand other essential system-level code occupied asubstantial part of the memory available in theTelosB motes that were used. As a result of thememory constraints, only a simple and lean securityimplementation could be attempted.

For encryption, the 128-bit AES algorithm wasused. AES is a block cipher which is obtained as aspecial case of the Rijndael algorithm. We ported anexisting C code for the algorithm into nesC for usein the motes.

For authentication and message integrity check,the CBC-MAC framework was used. In this, ablock cipher is used to calculate the messageauthentication code (MAC). Our implementationused AES itself as the block cipher.

Before deployment in the field, each sensor nodewas provided with the same secret key. Each messagewas encrypted using the secret key and the AESalgorithm. Authentication and message integritytags were calculated and transmitted along with themessage bits, using the AES-CBC-MAC framework.

Using a single secret key throughout the networkis risky, because a compromised node may end uprevealing the secret key and thereby jeopardizingcommunication over the entire network. To addressthis problem, periodic rekeying was implemented.The basic idea is that at regular intervals, each nodegenerates a new key for itself by executing a hashfunction with the old key as the argument. TheGrindahl hash function was used for this purpose.


Bayesian inference problemsare those in which thedecision maker is suppliedwith prior probabilisticinformation about thehypotheses.

Beamforming: A means bywhich signals transmitted bydifferent antennasphase-align at the receiver.

Markov decision process(MDP): Another term for acontrolled Markov chain,which is a Markov chain inwhich the transitionprobabilities depend on thecurrent state and the currentaction.


Loose time synchronization among the nodes wasassumed; this ensured that nodes executed the hashfunction at roughly the same intervals. However, toallow for occasional mismatches, a window of threekeys — past, present and future — was maintainedby each node.

A basic replay attack protection feature wasincorporated into the implementation. This reliedon including a timestamp with each message. Areplayed message would be caught by the oldtimestamp that it carried.

The implemented security features weredemonstrated on several occasions. An outsiderattacker not possessing the secret key and tryingto communicate is caught immediately by genuinesensors nodes in its vicinity, and an alarm messageis sent to the Base Station. Apart from periodicrekeying, the Base Station can command all nodesto switch to a new key at any time; this feature can beutilized to thwart a suspected adversary. Our currentwork is aimed at tackling the harder problem of an“inside adversary,” who has access to the secret keysand behaves in a malicious manner.

7. Event detection: other statisticalformulations

The simplest model for intrusion is one where sensorobservations are modeled as being sampled froma certain distribution prior to the intrusion, andas being sampled from a certain other distributionafter the intrusion. The observations at a particulartime instant are naturally spatially correlated,i.e., correlated across sensors, because they areobservations at the same sampling instant of thesame phenomenon (intruded or unintruded state).The goal is to detect the disruption (intrusion) assoon as it occurs, but not before. Given the stochasticnature of the system, one tries to minimise theexpected delay for detection after the intrusion,subject to a maximum tolerable false alarm rate.

7.1. Bayesian event detection in a small networkThe classical Bayesian sequential change detectionproblem [70] is formulated as follows. Thereis a random (discrete) time 01 at which thechange occurs. The prior information includes thedistribution of 01. All the sensors take periodicsamples of what they are designed to observe:temperature, strain, etc. In the small network setting,the change affects all the sensors in the same way, sothat conditioned on 01= t the observations at thesensors up to t−1 are assumed to be independentand identically distributed over time and acrossthe sensors, with probability distribution F0(·).Similarly, the observations are independent overtime and sensors from t onwards but with a different

probability distribution, F1(·). At each time k, thedetection procedure makes a decision as to whetherthe change has occurred at or before time k orto continue making measurements at time k+1.The optimal change detection procedure raisesan alarm at time τ (which depends only on theobservations made until then) such that the meandetection delay E

[(τ−01)

+]

is minimum over theclass of all stopping times τ whose probability offalse alarm PFA=P{τ < 01} does not exceed α, asuitably chosen small value, such as 0.01.

7.1.1. Physical layer fusionTypically, the observations made by the sensors

are quantised, coded, and then sent individuallyover the air to an information fusion and decisionunit. This assumes an ad hoc separation ofquantisation and transmission over the wirelesschannel. Moreover, each of these nodes contendsfor the common channel. In the classical approach,after receiving all observations, the fusion centrefuses these observations (for e.g., addition of loglikelihoods of individual observations) because theyare spatially correlated. Given this eventual additiveprocessing at the fusion centre, can superposition ofelectromagnetic waves provide this fusion? Suchan approach will not only do away with the need forchannel contention, but also perform a joint sourceand channel coding, a form of analogue coding.

In our recent work [79], we considered a “startopology” of a fusion centre surrounded by severalnodes with links from the nodes to the fusioncentre. The sensors first calculate the log-likelihoodratio of their observations, where the ratio is oflikelihoods under the changed distribution overthe likelihood under the no-change distribution.The sensor transmitter then amplitude modulatesa standard waveform with this value. It also appliesa phase correction based on channel knowledgefrom the base station to the node assuming thata TDD system is used and channel reciprocity.Superposition of the transmitted waveforms doesthe rest and computes the needed statistic fordecision making with the caveat that this statisticis embedded in receiver noise. This mechanism iscalled distributed beamforming, because waveformsalign at the receiving end. The fusion centre thenmakes a decision whether to stop or to continuesampling. It can also provide feedback to enableenergy savings.

The problem can be cast as a Markov decisionprocess. The decision at each stage is whether todraw a new set of samples or stop and declare anintrusion. When the decision is to draw a newsample, the choice of parameters of the affinetransmission optimise energy consumption. We


Partially observed MDP(POMDP): A type of Markovdecision problem in whichonly a noisy version of thestate is available to thedecision maker.

CUSUM algorithm: Aspecific sequential decisionalgorithm that accumulatesevidence in favour of thealternate hypothesis.


provide an explicit closed form solution for thecontrol at each stage, when the prior and posteriordistributions are Gaussian, but with different means.

The biggest challenge in this proposed schemeis the practicality of distributed beamforming.Clocks are required to be synchronised so thatcarrier phases and symbol ticks align. Engineersat the University of California, Santa Barbara haverecently embarked on an effort to build a prototypedistributed beamforming system [58].

7.1.2. Optimal sleep-wake cycling of sensor nodesIn the classical problem [70], the only tradeoff

being accounted for is that between detection delayand the probability of false alarm. In intrusiondetection applications using WSNs, the energy cost(i.e., sensing + computation + communicationcost) of using a sensor is a critical issue as sensors areenergy-limited devices, and an intrusion is typicallya rare-event. We consider a small extent network inwhich there are N sensor nodes. We are interestedin obtaining a detection procedure which, at everytime k, makes a decision as to whether the intrusionhas occurred or to continue making measurementsat time k+1 with Mk+1 ≤N sensors. Thus onlythese Mk+1 sensors will expend energy to providemeasurements at k+1, while the rest can stay ina low energy state. The energy-optimal detectionprocedure minimises the sum of detection delayand energy costs subject to the constraint on thefalse alarm probability.

The problem can be formulated as a PartiallyObservable Markov Decision Process (POMDP)with the information state at time k being 5k, theposterior probability of change having occurred at

Figure 20: Optimal sensor wake-up policy (number of sensors to beused as a function of the posterior probability) for F0=Normal(0,1),F1=Normal(1,1), sensing cost ls = 0.5, and false alarm cost lf = 100.0(this sets PFA to 0.04).

or before time k given the observations [64]. As isusual for such problems, a Bellman equation canbe written down. Some properties of the optimalpolicy can be obtained and numerical evaluationof the optimal policy can be done by value iteration.

From Figure 20 we see that when 5k is eitherlow or high, it is enough to keep one sensor awake,and when 5k is close to 0.5, we need to use moresensors. This can be interpreted as follows: whenthe posterior probability of change is low or high,the uncertainty about the occurrence of the event issmall, and when the posterior probability of changeis close to 0.5, the uncertainty about the event islarge. If the uncertainty is large, we need moresensors to resolve the uncertainty.

7.1.3. Detection over a network, and transient changedetection

In practice, the communication between thesensors and the fusion centre is facilitated bya wireless ad hoc network. The media accesscontrol (MAC) layer of the network introducesrandom delays in receiving the observations andhence, the observations may not be received inchronological order at the fusion centre. We studythe sequential change detection problem in thepresence of network delays in [67] and show thatthe optimal decision procedure requires additionalknowledge of the queueing state of the system.

Also, the change detection problem consideredso far assumes that the change persists for ever.However, in intrusion detection applications, thechange is transient (i.e., the change disappears aftera random time in the network). We formulate andstudy this problem in [66].

7.2. Non-Bayesian formulations: distributeddetection

7.2.1. The DualCUSUM AlgorithmWe have developed an algorithm, which we call

DualCUSUM, which is based on the well knownCUSUM algorithm [61]. In DualCUSUM a localdetection is made at each node by using the CUSUMalgorithm, and the local decisions at the nodes arefused at the fusion centre using another CUSUM.The following are some features of this algorithm.Details can be found in the related publications[9,37].

The local CUSUM threshold at each node ischosen appropriately to satisfy the overall systemfalse alarm probability target. Each sensor nodetransmits data (in fact just a bit) to the fusion nodeonly if it declares a change. As a change occursbut rarely in our intrusion application setting, inDualCUSUM a local node will need to transmit veryrarely. This leads to significant saving in energy used


Physical layer fusion: Themeans by which the signalprocessing step ofaccumulation of evidence atthe receiver is performed bythe superposition ofelectromagnetic signals.


in transmission. Node i transmits its bit at power Pi,which is a design parameter. All nodes can transmitsimultaneously. This saves MAC delays and leadsto overall reduction in detection delay at the fusionnode. For this physical layer fusion to be possibleone needs tight time synchronization of differentnodes (see Section 7.1.1).

Because of receiver noise, the fusion node doesnot know how many sensor nodes are transmittingin a slot. Thus, even if Pi and the sensor noisevariances are known to the fusion node, it cannotcompute the exact log-likelihood ratio needed forCUSUM. Thus, in DualCUSUM, at the fusion node,we use an appropriately designed density for the“alternate” hypothesis [9,37].

The overall performance (the probability of falsealarm and mean delay) of DualCUSUM depends onvarious parameters. We obtained the performanceof the algorithm for a given set of parameters. Wethen found the optimum parameters to obtain thebest performance. CUSUM requires knowledge ofthe node transmit powers, Pi, the receiver noisevariance, and the channel gains. Typically, these willnot be available. In such a scenario one can use anonparametric CUSUM or generalized likelihoodratio instead of the likelihood ratio. These variationsof DualCUSUM were also studied (see [37]).

7.2.2. Detection and localisation of an event in alarge WSN

We consider a large extent WSN, where an eventinfluences the observations of sensors only in thevicinity of where it occurs. Thus, in addition to eventdetection, we are also interested in identifying thelocation of the event. As we consider a large extentWSN, a centralised solution where a fusion centermakes a decision at each time k after receiving theobservations from all sensors, is not desirable. Wepropose distributed event detection and locationidentification procedures based on CUSUM [61]and analyse their optimality properties.

An event is modeled as occurring at a point in atwo-dimensional region of interest. Here we discussthe simple Boolean model where the post-changedistribution at any sensor within a radius of rs fromthe event is F1(·) and that for any sensor nodeoutside this disc is F0(·). A more general sensingmodel is considered in [65].

Figure 21 depicts the situation for the Booleanmodel. We partition the region of interest A in sucha way that each sub-region Ai in the partition iscovered by a unique set of sensors Ni (see Figure 21).Thus, we have a hypothesis testing problem foridentifying the location of the event where thehypothesis Hi corresponds to the set of sensors Ni

and the hypothesis H0 corresponds to change nothaving occurred so far.

Figure 21: An example of partitioning of Ain a wide-area WSN. The three sensor nodesdivide the overall region into regionsA1, . . . ,A7 such that region Ai is covered bya unique set of sensors Ni.

We study distributed procedures that detect andlocate an event (to any of the Ais) subject to a falsealarm and false isolation constraint. The false alarmconstraint considered is the average time to falsealarm at the region Ai (TFAi) defined as the averagenumber of samples taken under the hypothesis H0

to raise a global alarm and declare the hypothesisHi. The false isolation constraint considered is theaverage time to false isolation TFIij defined as theaverage number of samples taken to raise a globalalarm and declare the hypothesis Hj when the eventhas not occurred in Aj but has occurred in Ai. Thesupremum average detection delay of a procedureτ, SADD(τ) is defined as the worst-case averagenumber of samples taken under any hypothesisHi, i= 1,2, . . . ,N , to raise a global alarm.

Each node runs a local decision rule, and theglobal decision rule is to stop and raise an alarm assoon as all the sensors in any set of sensors Ni havea local decision 1, and we isolate the location ofthe event to the region Ai. We propose three localdecision rules: ALL (see [54]), MAX (see [74]), andHALL (hysteresis modified ALL), all of which arebased on running the CUSUM procedure at eachnode.

We show in [65] that the procedures ALL andHALL are asymptotically minimax delay (SADD)optimal as the mean time between false alarms (andfalse isolations) goes to infinity. We compare theperformance with an optimal centralised proceduredue to Nikiforov [60], and find that the distributedprocedures ALL and HALL achieve almost the sameperformance as that of Nikiforov’s centralisedalgorithm.


Accelerometer: Ameasurement device thatbehaves like a mass on aspring. The displacement ofthe proof mass isproportional to the externalacceleration.

Noise floor: The level of allunwanted signals fromwithin and outside themeasurement system.


8. Development of a high-resolutionvibration sensor

We present here an overview of the work done onthe development of a single-axis micromachined in-plane accelerometer for use in detecting intrusionby sensing the ground vibrations. Detection ofintrusion caused by human foot-steps and vehiculartraffic in rough terrain requires sensitivity froma vibration sensor of the order of 10−6 to 10−3

g over a few kHz frequency range (Mazarakisand Avaritsiotis [53]; Ekimov and Sabatier [23]).Commercially available vibration sensors that candetect millionth of the acceleration due to gravity(the so-called micro-g accelerometers) are veryexpensive (e.g., about Rs. 30,000 to 1,00,000 persensor) and hence they are not economically viablefor wireless sensor networks. So, the objective ofthis work was to design and develop a low cost,high resolution, highly sensitive, small-sized, power-efficient vibration sensor that works over a largebandwidth. A summary of the work done duringthe course of this project is described here.

8.1. Initial trials with a commercial accelerometerIn order to assess the feasibility of humanfootstep detection, we conducted some trials witha commercial accelerometer (LIS3LO2AS4; STMicroelectronics (http://www.st.com/stonline/).The chip along with the evaluation board is shownin Figure 22(a). It is a three-axis accelerometer witha sensitivity of 500 mV/g. Its specification states thatits noise-floor is 50 µg/rootHz. However, in our

trials we could not see a resolution better than afew mg when it was tested on a shaker table (LSDshaker). This may be because of the ambient noiseas well as due to the mounting. This, as statedearlier, is a concern with vibration sensors as theyare sensitive to the way they are mounted. To checkfor consistency, several evaluation boards as wellas mountings were made in-house. This is shown inFigure 22(b). Furthermore, we attached them to ametal peg firmly so that the sensor platform can befixed into the ground as shown in Figure 23(a–b). Apower supply unit and an oscilloscope were kept onthe side to measure the performance. It was noticedthat the sensor was able to pick up human footstepswithin a distance of about 1 m away from the sensor.Shown in the right panel of Figure 23 is a clear signalof about 20 mV due to a foot-thump at a distanceof 1 m. It can be noticed that there was a noise ofabout 3 mV. This trial confirms that footsteps can bedetected. For long-range detection, we need higherresolution than what was seen in this trial.

8.2. Initial design and system level simulationThe work began with a focus on an in-planecapacitive accelerometer. The design of themechanical sensor element and the capacitance-sensing circuit began simultaneously. The sensorelement design assumed that a capacitance-changeof one part per million could be detected by thecircuit. This set the goal for designing the sensitivityof the accelerometer and the minimum detectableacceleration. The circuit design assumed that thesensor element has a base capacitance of about 1 pF.

Figure 22: A commercial accelerometer from ST Microelectronics (LIS3L02AS4) with the specifications:sensitivity of 0.5 V/g, noise-floor of 50 µg/rootHz, resonance frequency of 1.5 kHz, and a cross-axissensitivity of about 2%. (a) Sensor chip with the in-house built evaluation board and mounting; (b)schematic of the mounting and the sensor chip.


Displacement-amplifyingCompliant Mechanism(DaCM): A monolithicjointless lever that amplifiesdisplacements using elasticdeformations.

Cross-axis sensitivity: Thesensitivity of the device in anaxis perpendicular to theintended axis.

Bandwidth of a sensor: Thenumber of times a reliablesensor reading can be takenper second.


Figure 23: Testing of the accelerometer mounted in the ground. (a) The vibration sensor, the powersupply unit, and the oscilloscope; (b) close-up view of the sensor; The scope picture on the right showsthe captured signal for a foot-thump in the vicinity of the mounted sensor; 3 mV noise-floor and 20mV signal.

(a) (b)

An in-plane accelerometer was chosen ratherthan an out-of-plane accelerometer becausea distinguishing feature of our approach wasto enhance sensitivity through mechanicalamplification and that is more amenable inthe plane than the out of the plane motion.The aim here is to enhance the sensitivity andresolution of the accelerometer. This can be done byimproving the capacitance-extraction circuit andsignal conditioning and by amplifying the signalmechanically in the sensor element itself. With apreliminary design of the sensor element (Krishnan[42]) and the capacitance-extraction circuit (Hegdeet al. [32]), it was noticed in the system-level analysis(see Figure 24) that the noise is, in general, morein the electronics than in the mechanical sensorelement. Therefore, even though we can increasethe sensitivity in both the mechanical and electronicparts of the sensor system, it makes more sense toincrease the sensitivity in the one with less noise.This way, we do not amplify the overall noise as weamplify the signal. So, it was decided that we usemechanical amplification of the acceleration (i.e.,vibration) signal to get enhanced signal-to-noiseratio.

Mechanical signal amplification is achievedwith a Displacement-amplifying CompliantMechanism (DaCM) (Krishnan and Ananthasuresh[43]). Figure 25 shows a DaCM appended toa conventional accelerometer. A conventionalaccelerometer consists of only a proof-masssuspended by its own folded-beam suspension. Theinput end of the DaCM is appended to the proof-mass as shown in the top portion of Figure 25. Theoutput end of the DaCM displaces much more thanits input end and hence the capacitance-sensing

comb is located at the top most part of Figure 25. ADaCM is possible for out-of-plane motion too butthe required microfabrication process will be muchtoo complicated. Hence, in-plane accelerometersare considered in this project. To demonstratethe feasibility of this concept, a meso-size (mmto cm) spring-steel accelerometer was made andwas interfaced with a PCB implementation of thecapacitance extraction circuit. The details of thiswere reported in [11].

8.3. Micromachined chip and ASICThere are many trade-offs in accelerometer design:between sensitivity and bandwidth; betweensensitivity and range; between main and cross-axis sensitivities, etc. Finally, the resolution, whichdepends on noise in the mechanical and electronicsparts, also has considerable trade-off with otherperformance parameters (Krishnan et al.[44]). Ingeneral, thick proof-mass, low stiffness of thesuspension in the main axis, closed-loop operation,and differential capacitance arrangement help.We used optimization techniques to obtain threecompeting designs as indicated in Table 2 (Krishnan[42]). The bandwidth of all of them was around 1kHz. It can be noticed in Table 2 that what givesthe best resolution does not have the best cross-axissensitivity and range. These were designed witha silicon-on-insulator (SOI) process in mind. Weassumed a proof-mass thickness of 250 µm andthe suspension thickness of 25 µm. We kept theminimum width of the narrowest beam at 5 µm.The layout of one of the accelerometers is shown inFigure 26 with its components marked.

Due to lack of access to an SOI process thatcan give thick proof-mass (see Figure 28, left



Figure 24: System-level simulation of the accelerometer performed in Simulink in Matlab.

panel), we reduced the proof-mass to the samethickness as that of the suspension and DaCM(see Figure 28, centre panel). This reduced theresolution to mg rather than 10s of µg. Two designsof milli-g accelerometers using SOIMUMPs (Silicon-on-insulator Multi-user MEMS Processes) werefabricated in MEMSCAP, USA, a microsystemfoundry. The process-flow of SOIMUMPs limitsthe thickness of the proof-mass to 25 µm [56] thusreducing the resolution and sensitivity of the device.A photograph of the SOIMUMPs die containingthe accelerometer is shown in Figure 28 (rightpanel). The size of the die was 1 cm×1 cm whereasthe size of the accelerometer on the die was 6.89mm×6.07 mm. The sense-gap between the fingersof the capacitance-sensing combs was 10 µm. Aclose-up view of the fabricated device is shown inFigure 27. The base capacitance of the sensors was15.5 pF when the substrate was electrically floating.It was 2.6 pF when the substrate was grounded.All the measurements were performed at a supplyfrequency of 1 MHz and a DC bias of 1 V with anAC level of 5 mV. The simulations showed the valuesto be 12.08 pF and 2.08 pF, respectively, a reasonableagreement with the measured capacitances.

Table 2: Simulated performance parameters of three competing designs for anSOI-process

Design Sensitivity Cross-axis Resolution (µg) Range (mg)(V/mg) sensitivity (%)

1 0.125 0.030 40 402 0.445 0.010 20 63 0.070 0.005 70 70

Figure 25: Displacement-amplificationcompliant mechanism appended to anaccelerometer.

The die-on-board packaging technique wasadopted to package the SOIMUMPs die containingthe milli-g accelerometer. A custom designed gold-plated PCB was used as a base onto which theSOIMUMPs die was pasted using conductive silverepoxy. In order to bond the contact pads on theaccelerometer to the gold-plated bonding siteson the PCB, 1 mill Aluminum wire was used. Atransparent acrylic cap was used to seal the wirebonded die hermetically to protect it from possible



Figure 26: Schematic of the mechanical sensing element with the displacement-amplifier, proof-masssuspension, suspension on the sensing side, and comb arrangement for capacitance detection andfeedback stabilization.

Figure 27: A close-up view of the displacement-amplifying compliant mechanism (DaCM) as fabricatedin an SOIMUMPs run.



Figure 28: Left and center panels: Solid models of the high-resolution accelerometer. Right panel: The fabricated SOIMUMPs dieshowing its various parts.

Figure 29: (a) The system-on-board packaged SOIMUMPs die, (b) The die on the PCB afterwire-bonding.

Figure 30: Block diagram of signal conditioning circuit prototyped on0.35 µm technology.

contamination. The packaged die is shown in theFigure 29.

The packaged SOIMUMPs accelerometer withdifferential comb-drive senses external accelerationby changing the base capacitance of the comb-drive.A signal conditioning circuit called a CapacitanceSensor is required to convert the change incapacitance to change in voltage. Two differentcapacitance sensors were used for integration tothe packaged SOIMUMPs accelerometer: (1) AMS3110 Universal Capacitive Readout IC [57], (2)A Universal Capacitance Sensor ASIC [32], anddeveloped as part of this project. The inputs tothese Capacitance Sensors are the C+ and C− portsof the differential comb-drive capacitance. Theaccelerometer integrated to the Capacitance sensingboard was placed on a LDS shaker table whichcan shake at a given acceleration. A commercially


Noise figure: A measure ofthe amount of noise that thereceiver adds to the radiosignal coming from theantenna.

Signal to noise ratio (SNR):The amount of signal powerper unit noise power, usuallymeasured in decibels.

Bit error rate (BER): Thefraction of bits, amongst allthose transmitted, that arereceived in error.


available 3-axis analog milli-g accelerometer, madeby ST Microelectronics (Model: LIS3L02AS4), wasused as a standard accelerometer for calibratingthe SOIMUMPs accelerometer and controlling theshaker table.

The capacitance to analog voltage conversionASIC developed on 0.35 µm technology platformcan sense single ended or differential capacitancechange up to sub-femto range. The ASIC based onContinuous Time Voltage sensing [51] methodologyis robust to the effect of parasitics largely present inthe system and exhibits a static sensitivity of 30 mV/fF for a gain of 10. The signal conditioning has amodulation and a coherent demodulation modulewith programmable features such as the bandwidthand the sensor offset correction to customize theASIC for a specific MEMS transducer. The signalconditioning block diagram is as shown in Figure 30.The packaged milli-g accelerometer was interfacedwith the IISc ASIC and then tested on a shaker table.The mounting arrangement is shown in the leftand centre panels of Figure 31, and the dynamicresponse of the accelerometer is shown in the rightpanel. Due to packaging, parasitics, and noise, theminimum acceleration measured was 30 mg [40].Efforts are underway to improve the performanceand also fabricate the device with a thick proof-massto achieve micro-g resolution.

9. Towards low power WSN motes9.1. A low power adaptive radioIn a typical mote used in a wireless sensor network,the radio transceiver module is the dominant powerconsumer. For example, in Texas Instrument’schipset for a mote, the radio current is about 20 mA

each for both receiver and transmitter, while that forthe rest of the micro-controller functions is less than5 mA [14,1]. We will focus on techniques we havedeveloped to reduce the power of the receiver (seealso Section 10.1). We will first give a brief overviewof a typical receiver architecture and then introducetwo main ideas we have explored to reduce thereceiver power, followed by some simulation andmeasurement results.

The key parameters in the design of the receiverare the noise figure and linearity. The noise figureindicates how much noise the receiver adds tothe RF signal obtained from the antenna. Therequirement is that the added noise, and hencethe consequent degradation in signal to noise ratio(SNR) be sufficiently small such that the target biterror rate (BER) is still achieved. The added noise inthe analog section can be reduced by using largercurrents to bias the circuits. In the digital section, byincreasing bit-width and sampling rate, the addeddigital noise can be reduced, albeit at an increasedpower consumption. From the desired BER, onethen derives the maximum noise figure and theminimum linearity and proceeds to design thereceiver by burning the requisite amount of currentto achieve this. However such a methodology leadsto a very conservative design for the situationswhen the received signal strength is much larger orthere is not much interference in the neighbouringfrequency bands. A key contribution of our work isto develop a technique to make the receiver powerscalable, that is, it reduces its power consumptionif the received signal strength and interference arebetter than expected [21].

Figure 31: Left panel: DaCM accelerometer interfaced onto the ASIC; centre panel: Integrated system mounted on the shakertable to sense milli-g acceleration; right panel: Dynamic Response of the integrated system developed at IISc for an inputacceleration of 400 mg at 20 Hz.


Sampling rate: The rate atwhich the received analogsignal is sampled anddiscretised in time.

Digital bit width: The numberof bits used to represent thesampled analog value.


Figure 32: Power scalable digital backend design.

9.1.1. Power scalable receiver designA variable gain amplifier (VGA) is a commonly

used block in any receiver analog front end. Thevariable gain amplifer adjusts the gain in thesignal path such that a full scale voltage can bemaintained at the inputs of the ADC. Thus, whenthe received signal strength is large, gain is reducedand vice versa. However, while this gain adjustmentis fairly common, we propose to adjust the powerconsumption of the entire receiver in converse tothe received signal strength. Furthermore, we alsolook at the strength of the interfering signals in theneighbouring band to determine how to adjust theoverall power consumption. Figure 32 shows ourpower scalable receiver design for the digital backend. The main additional components for makingthe receiver power scalable are: a) tuning knobs forthe ADC and the digital datapath to dial down thepower consumption as required b) an adaptivitycontrol unit consisting of a signal and interferenceestimation module and a scalable filter design.

Scalable power control for the ADC and the digitaldata path

The two main knobs for trading off power withperformance are the sampling rate and the digitalbit width. The sampling rate is adjusted by adjusting

the frequency of the clock being used by thesesections. The maximum rate is 30 Msps and theminimum is 2 Msps. The higher sampling rateallows for better averaging of the noise and henceresults in better BER, but at the cost of increaseddynamic power dissipation. The bit resolution isadjusted by selectively masking the bits at the inputto the digital data path (i.e. the bits from the ADCare ANDed and shifted down). For example, themaximum bit resolution is 8-bits and the entire datapath is designed for this. However, under certainbenign conditions, even a 1-bit data path suffices. Toaccommodate this, the lower 7-bits from the ADCare masked and the 8th bit is right shifted all theway to the first bit and fed into the data path. Thesame is done for all the other input signals in thedatapath — like the input from the digital oscillator.This ensures that the switching activity is restrictedto only the lower order bits, while the higher orderbits remain inactive, thus saving dynamic power.

Adaptivity control unitThe adaptivity control unit measures the

interference and signal power levels during thepreamble of the packet in a non-coherent manner.The measurement estimates are then used todetermine the operating parameters for the receiver


dBm: The decibel value of themeasured power inmilliwatts.


for receiving the data symbols of the packetfollowing the preamble.

The two key components of the adaptivitycontrol unit are the interference and signal powerestimation module and the lookup table whichdetermines the receiver’s operating parameters. The802.15.4. standard specifies the BER requirementsfor the case when the signal amplitude is the weakestat – 85 dBm, the adjacent channel has a power of– 85 dBm and the alternate channel has a powerof – 55 dBm. The baseline sampling rate of 30Msps and resolution of 8-bits for the ADC anddigital data path ensures that the target BER is metunder these conditions. When there is not muchinterference or the received signal is stronger, wedetect this condition and dial down the samplingrate and resolution down up to 2 Msps and 1-bit. Wedetermine the energy of the adjacent and alternatebands by downconverting them and measuringthe power over a duration of 4 chips. Each chipis a half-sine waveform modulated as±1. Sixteenchips each for the in-phase (I) and quadrature(Q) components forms one 4-bit symbol in the802.15.4 standard. Signal power is also similarlymeasured. The same hardware is reused for allthese measurements to save on gate count. Themeasured interference power is quantized to 1-bit each for adjacent and alternate channels anda 6-bit quantization is used for representing thesignal power. These 8-bits are then used to look upinto a precharacterized table, which tells the rightsampling rate and quantization to use for receivingthe signal under these conditions, while meetingBER. This lookup table is characterized a prioriusing MATLAB.

Scalable filter designWhen the sampling rate is adjusted to be less

than its maximum values, the filter in the data pathgets affected and special attention needs to be paidto ensure that the overall transfer function is stillmaintained, invariant to the sampling/operating

Table 3: Mica2 platform current draw, measured with a 3V power supply.

Device/Mode Current (mA) Device/Mode Current (mA)

CPU Radio

Active 8.0 Receiving 7

Idle 3.2 Transmitting(-20 dBm) 3.7

ADC Acquire 1.0 Tx(-8 dBm) 6.5

Extended Standby 0.223 Tx(0 dBm) 8.5

Standby 0.216 Tx(10 dBm) 21.5

Power-save 0.110 Sensors

Power-down 0.103 Typical board 0.7

rate. The filter used is a matched filter, and isimplemented by correlating the received sequencewith the samples of the half-sine modulating pulsefor the chip. This pulse was generated by a CORDIC[4] algorithm. When the sampling rate is 30 Msps,all the filter taps are used. Since the filter coefficientsare symmetric about the central value, the numberof multipliers is reduced to half. When the samplingrate is reduced from 30 Msps to say 3 Msps, thenonly three taps are kept active, with zeros fed asinputs to the other taps. This scales the dynamicpower dissipation accordingly.

9.1.2. Current statusThe design of the receiver has been implemented

in the Verilog Hardware Description Language(HDL) and has been synthesized for the UMC130 nm process technology. The power for thedifferent operating conditions was estimated usingSynopsys Power Compiler. Savings of 85% in powerwere obtained when the receiver is operated at itsminimal configuration (for benign condition) ascompared to the worst case setting for the harshestinput conditions.

Our main motivation is to take advantage ofbenign channel conditions and dial down thepower consumed by the receiver. We achieve this byincorporating an interference and signal estimatorblock which measures the signal and interferencelevels during the preamble. A precharacterizedlookup table is then consulted to set the samplingrate and bit resolution of the digital data path tobe as low as possible for the rest of the packetreception. We also implement the timing recoveryalgorithms in a streaming fashion, to minimize thepower dissipation associated with storage. Withthese techniques, a receiver in a UMC 130nmprocess shows upto 85% power reduction betweenits maximum and minimum power settings, thusshowing power scalability.

10. Mote system software10.1. Compiler assisted energy management for

sensor network nodesThe major sources of power consumption in asensor network node are the transmission unitand the computing unit. Table 3 adapted from thestudy done in [71] shows the typical current drawnin various units in a sensor network node. FromTable 3 we observe that indeed the power consumedby a radio is significant, and power required forcomputation, though lower, is quite significant.Given the ability to operate the processor in bothactive and idle mode these numbers do not presentthe complete picture. The computational demandsof the application determine the CPU’s actual


Dynamic voltage scaling(DVS): A powermanagement technique thatincreases or decreases thevoltage of a componentsystem as per the need foruse of the component.

Cache: A component thatstores data so that futurerequests for the same datacan be served faster.

Scratchpad: A high-speedinternal memory fortemporarily storingintermediate calculations.

Embedded system: Acomputing system,embedded in a larger system,and designed for specialisedfunctions with real-timecomputing constraints.


activity and radio usage. In [71], the authors alsopresent a detailed breakdown of energy consumedby different components for a variety of applications.The results suggest that CPU power ranges from 28%to 86% of the total power consumed and roughly50% on average. Applications like data filtering,encryption and decryption (required in networksused for critical tasks such as military surveillance)and more efficient communication protocols canshift the activity from the radio to the CPU. Hence,saving CPU energy will also play a crucial role.

Researchers in the sensor network domainhave focused more on optimizing energy in thecommunication system than in the computingsystem. Researchers have tackled the problem ofenergy optimization in communication system,from various aspects. Some proposals try to makeuse of power aware radios [73,76], while otheruse more efficient modulation schemes [69] andenergy aware sensor network protocols [33,63,78].Researchers have also suggested using energy awaresoftware (see Section 9). For example, implementingdynamic power management through operatingsystem [10,35]. Recently, there has been efforts forminimizing energy consumption in computingsystems as well. In this respect researchers haveproposed low power architecture design for sensornetwork node processors [46,22,34,59]. Thesedesigns suggest use of specialized hardware unitsfor performing tasks in a sensor network node.

An important requirement of the energymanagement techniques used for sensor networksis that the extra computation and communicationintroduced by the energy optimization schemesmust be low. Otherwise, the energy required toperform the optimization schemes may outweighthe benefits. Conventional techniques in sensornetwork domain try to minimize communicationenergy, as conventional communication linksconsume a significant amount of energy. However,recent developments in self-powered MEMS-basedRF communication devices [49] can lead to sensornetworks where the communication link is entirelyself-powered. Hence, computational energy willbecome even more important.

We have focused on compiler assisted energyoptimization techniques for sensor network nodeprocessors. To the best of our knowledge little workhas been done in compiler related optimizationtechniques for sensor networks [46]. In compilerassisted energy optimization strategies majordecisions are taken offline. During runtime, basedon the current system state, the appropriate decisionis invoked. On the other hand, for OS leveltechniques, all the decisions are to be made atruntime, incurring energy cost on already energyconstrained nodes. Our main contribution areoutlined as follows.

Compiler assisted dynamic voltage scalingDifferent sensing applications will have different

processing requirements in the active state. Havinga processor with a fixed throughput is necessarilypower inefficient. Having a custom digital signalprocessor for every sensing application is notfeasible both in terms of the cost and timeoverhead. However, energy savings can still beobtained by the use of dynamic voltage scaling.We have implemented and evaluated a compilerdirected dynamic voltage scheme for sensornetwork nodes. The implemented technique leadsto an average energy savings of 29.04% at aperformance slow down of 2.88%. On average26.99% of improvements in energy-delay productare observed.

Compiler assisted scratchpad memory allocationMemories are a major consumer of energy

in a system. They also limit the performanceof the processors. In general purpose processorscaches have been used to reduce the memoryaccess time. Caches are power hungry and alsointroduce non-determinism in the system. Hence,embedded systems use more energy efficientscratchpad memories. In this project we show thebenefits of using scratchpad memory in sensornetwork node processors. We have implemented acompiler assisted scratchpad allocation algorithm.On an average 50% of energy savings and 52%performance improvements are obtained with a512 byte of scratchpad memory. We also found thatthe scratchpad size for a system should be chosencarefully, as it may have a negative impact on energysavings.

We also studied the effect of the presence ofscratchpad memory on the dynamic voltage scaling.Our results show that the presence of scratchpadmemory in the system limits the opportunitiesfor dynamic voltage scaling. We also performeda comparison among all the three optimizationcases (DVS, Scratchpad, Scratchpad and DVStogether) normalized to the no-optimization case.Our comparison results show that benefits of ascratchpad memory in a system are comparable tothe benefits of scratchpad and DVS applied togetherand improve further with an increase in scratchpadsize.

10.2. Thread-based sensor node OS with limitedstack

Sensor nodes exhibit characteristics of bothembedded systems and general-purpose systems.A sensor node operating system is similar toan embedded operating system, but unlike ina typical embedded operating system, sensor


Multithreading: Aprogramming paradigm thatallows multiple threads ofcomputation to coexist andshare the resources assignedto a process.

Preemptive scheduling: Ascheduling paradigm inwhich a task may beinterrupted in order to serveanother task of higherpriority.

Stack: A segment of memoryassociated with a process,and with last-in-first-outaccess.


network applications may not be real time, andare constrained by limited memory, computationalpower and energy available in each sensor node.The computational power and program memoryis restricted to minimize power consumption asthe motes run on battery power. Therefore, it isimportant to design an operating system for sensornodes that attempt to minimize the use of memoryallocated at runtime, i.e., dynamic memory suchas stack. Other basic requirements of wireless sensornetwork OS are the following.

Reliable computation and communicationIn order to ensure reliable computation, it is

necessary to detect faults such as stack overflow.If dynamic memory is not used for stack, it maypossibly be easier to detect stack overflow as theposition of stacks would be fixed. It is also importantto ensure end-to-end reliable communication forthe proper functioning of the WSNs.

User-friendly programming environmentThere are two aspects of user-friendly

programming environment, viz., (i) programmingusing well known programming language, such asC, and using well known programming paradigmlike thread paradigm, and (ii) a good programminginterface in the programming paradigm used. Mostof the OSs for WSNs [68] have been written in C,but they may not use thread paradigm becauseof the use of a stack for every thread. A few otherprogramming languages such as nesC [18], galsC[26], etc., have also been proposed for writing OSsfor WSNs. In TinyOS [48], both the OS and theapplications have been written in nesC which isharder to learn, and therefore adds to the additionalproblem of developing and debugging the system.

Preemptive multiprogrammingEven though the applications for WSNs may

not be real time, some of the distributed actionsof the OS for WSNs may have to be synchronized,and therefore, it is desirable to support preemptivemultiprogramming along with thread programmingparadigm. It is also necessary that the threadprogramming paradigm has a good programminginterface.

By considering the above requirements, we haveproposed a thread-based execution model that usesonly a fixed number of stacks. In this executionmodel, the number of stacks at each priority levelis fixed. It minimizes the stack requirement formulti-threading environment and at the sametime provides ease of programming. We also takeadvantage of the fixed position of stacks to detectstack overflow, and therefore increase the reliabilityof the OS for sensor nodes. We give a separateimplementation of this thread model in the ContikiOS [19] without using any part of protothreadimplementation in the Contiki OS.

10.2.1. Proposed model and implementationDue to the presence of long running tasks and

some basic real time requirements of sensor nodeOS, such an OS should allow preemption andpriority-based scheduling. We have chosen Contikias our development platform, as it provides manyfeatures like dynamic loading and unloading ofprograms and services. Keeping the basic eventdriven model of Contiki kernel as it is, we embeddedwithin it our thread-based model. In our model,the number of priority levels (MAX LEVELS) (i.e.,priority queues) is fixed. Depending upon thememory available, we have also fixed the number ofstacks for each priority level (STACKS PER LEVEL),and from this pool of stacks, a stack is allocated to athread in a priority level before it gets scheduled.Therefore, at any point of time, the number ofthreads in Ready state is upper bounded by themaximum number of stacks.

Total stack space required by the system isbounded by treating it as a limited resource. Stackmust be allocated to a thread before moving it toLT READY state as shown in Figure 33. The stack willbe released explicitly either by the thread blockingfor an event or by the system when the thread exits.It is required that the thread in a priority level eitheralways hold the stack or always release the stack,while waiting for an event. If a priority level requiresthe threads to hold the stack while waiting for anevent, then the priority is said to be a stack holdingpriority; otherwise, the priority is said to be a stackreleasing priority.

Initially, when a thread is created (LT CREATEDstate), the stack is not allocated to it. It willbe moved from LT CREATED state to LT READYstate only when stack is allocated for this thread;otherwise it is moved to LT WAITING STACK state.If a thread is in LT RUNNING state with stackholding priority, it can wait for an event holdingthe stack. Similarly, if a thread is in LT RUNNINGstate with stack releasing priority, it can wait foran event releasing the stack. If the thread is inLT BLOCKED RELEASE state, and the event occurs,the thread is moved from LT BLOCKED RELEASEstate to LT WAITING STACK state or to LT READYstate depending upon the availability of the stack.Another possibility is that when a thread wants tochange its priority either to stack holding priorityor to stack releasing priority, it directly moves fromLT RUNNING state to LT WAITING STACK state.When a thread releases stack, and later on it obtainsa stack again, the location of the new stack may bedifferent from that of the previous stack allocatedto it as we can have multiple stacks at each prioritylevel.

A typical application of release stack and wait foran event can be a thread that waits for encrypting



Figure 33: The new state transition diagram.

Figure 34: Snapshot of threads active in system for one stack per level.

a message. Once the encryption of the message isdone, it can choose to release the stack, and wait forthe next message to be encrypted. The advantageof having the option of releasing stack and waitingis that we do not need to waste resources on theoverhead of creating context of threads again andagain for repetitive tasks.

10.2.2. Scheduling policyOur proposed model uses the following

scheduling policy.

• Fixed priority preemptive scheduling policyis used.

• FIFO is used for threads having the samepriority.

• If a thread with stack releasing priority movesto a stack holding priority, it does not holdany resources.

• If a thread with stack holding priority movesto a stack releasing priority, it first releasesthe stack, but can continue to hold otherresources.

The implication of the above scheduling policyis the following.



• A low priority thread is scheduled only ifno higher priority thread is present in readystate.

• The currently running thread is preempted ifanother thread with priority strictly greaterthan that of the currently executing thread ismoved to ready state.

The number of priority levels as shown inFigure 34 are fixed but can be changed at compile-time. Priority levels cannot be added or removedat runtime to reduce overhead of cleaning priorityqueues and managing priority conflict. Using FIFOscheduling policy among the threads with the samepriority, the number of preemption needed will bemuch lesser compared to that needed by a roundrobin scheduling policy. As the number of contextswitches decreases, the energy consumption will alsodecrease. Using this scheduling policy, we can takemaximum advantage of dynamic voltage scaling,as the threads in a priority level with non-real timerequirements can be executed at a slower pace.

There is no possibility of deadlock due tocontention for stack in the above scheduling scheme.This is because a thread in stack holding prioritylevel does not hold a resource while waiting for astack. Therefore, no other threads will wait for it,while it waits for stack. A thread in stack releasingpriority level can hold resources while waiting forstack, but it is always assured of getting a stack asother threads in the same priority level with readyor running state will continue to be executed untilthey release their stack.

Thread 0 is the default thread for system withlowest priority, and it monitors the system. Ifthere are no other threads to be scheduled, itswitches to power saving mode if supported bythe microprocessor.

Our thread model has been implementedwithout using any part of protothreadimplementation, and applications using our threadmodel do not need to use protothreads. However,protothreads can be used together with our threadsas higher priority threads using the stack space ofthe event driven kernel stack.

11. ConclusionWe have reported on the outcomes of a researchand demonstration project on human intrusiondetection in a large secure space using an adhoc wireless sensor network. As one would haveexpected, there were aspects of our work specific tothis application, and other general aspects relevantto a wider set of applications. After comparingacoustic, vibration, and passive infrared (PIR))sensors, we found PIR sensors to be the most

effective for our application. A PIR “trip-wire” wasdeveloped by wirelessly networking several mote“platforms”, each with three off-the-shelf quad-pixelPIR devices. Each such platform is able to cover a360◦ field of vision, and is equipped with a supportvector machine derived classification algorithm,which is trained to distinguish human intrusionfrom vegetation “clutter”. Several such parallel trip-wires help reduce the probability of false alarm.Such a detection system is connected to the operatorstation by a multihop wireless ad hoc network oflow power relay nodes. The network was designedto be self-organising, self-routing, and self-healing.Several security features have also been includedin our design. In addition to the demonstrablesystem, which we call SmartDetect, the projecthas yielded new basic research in the areasof self-healing geographical routing, distributedsequential algorithms for change detection inrandom processes, a MEMS accelerometer witha capacitive output, and a femto-Farad capacitancemeasurement circuit, that can be used for footstepdetection, the design of and some components of amicro-radio that can adjust its power consumptiondepending on the signal-to-noise ratio of the signalit receives, and energy efficient embedded systemssoftware.

Considerable engineering work remains tobe done to convert SmartDetect into a robustdeployable system. We propose to follow up onsome of the scientifically interesting challenges in apossible second phase of the project. One of the areaswhere wireless sensor networking can be expectedto be of significant commercial importance is inindustrial sensor networking, where 100s of sensorsin an industrial setting need to be interconnectedwirelessly. This problem throws up the challengeof designing low cost and low power wireless meshnetworks with predictable performance in termsof delay and reliability. This is another challengethat we are currently pursuing in our work.

AcknowledgementWe would like to thank Prof. V.K. Aatre (visitingprofessor ECE, IISc, and former Scientific Advisorto the Defence Minister) for his support andencouragement, and Dr. S.V. Sankaran, DirectorER&IPR, DRDO, for his support in the manyadministrative matters that we encountered duringexecution of the project. The DRDO lab weinteracted with in the course of this project wasthe Centre for Artificial Intelligence and Robotics(CAIR). We would like to thank Dr. N. Sitaram,former Director of CAIR, and Dr. V.S. Mahalingam,current Director of CAIR. Our principal pointof contact at CAIR was Dr. N. Ramamurthy. We



thank him and his team of engineers for meeting usfrequently, and providing very constructive feedback.Prof. A. Arora from Ohio State University, was avisiting faculty member to our project, and provideda wealth of useful suggestions. Thanks are also dueto Dr. A. V. Krishna from PESIT, Bangalore, forintroducing Support Vector Machines to this team.

Received 14 May 2010; revised 15 October 2010.

References1. IEEE 802.15.4. Part 15.4 : Wireless medium access control

(MAC) and physical layer (PHY) specifications for low-ratewireless personal area networks (LR-WPANs). October2003.

2. P. Agrawal and J. Kuri. A Key Distribution Scheme withoutDeployment Knowledge. In Proceedings of the InternationalConference on Ultra-Modern Telecommunications (ICUMT),2009.

3. J. Altman. Acoustic and seismic signals of heavy militaryvehicles for co-operative verification. Journal of Sound andVibration, 273:713–740, 2004.

4. Ray Andraka. A survey of CORDIC algorithms forFPGA based computers. In FPGA ’98: Proceedings of the1998 ACM/SIGDA sixth international symposium on Fieldprogrammable gate arrays, pages 191–200, New York, NY,USA, 1998. ACM Press.

5. Applied MEMS, Inc., www.appliedmems.com.6. Anish Arora, P. Dutta, S. Bapat, V. Kulathumani, H. Zhang

an V. Naik, V. Mittal, H. Cao, M. Demirbas, M. Gouda, Y. Choi,T. Herman, S. Kulkarni, U. Arumugam, M. Nesterenko, A. Vora,and M. Miyashita. A line in the sand: a wireless sensor networkfor target detection, classification and tracking. Journal ofComputer Networks, 2004.

7. Anish Arora, Emre Ertin, Prasun Sinha, Ted Herman, NishankTrivedi, Mikhail Nesterenko, Mohamed Gouda, David Culler,Mike Grimmer, and Ken Parker. Exscal: Elements of anExtreme Scale Wireless Sensor Network. Technical report,The Ohio State University, University of Iowa, Kent StateUniversity, Michigan State University, The University of Texasat Austin, and University of California at Berkeley, 2004.

8. Anish Arora, Rajiv Ramnath, Emre Ertin, Prasun Sinha, SandipBapat, Vinayak Naik, Vinod Kulathumani, Hongwei Zhang,Hui Cao, Mukundan Sridharan, Santosh Kumar, Nick Seddon,Chris Anderson, Ted Herman, Nishank Trivedi, Chen Zhang,Mikhail Nesterenko, Romil Shah, Sandeep Kulkarni, MaheshAramugam, Limin Wang, Mohamed Gouda, Young ri Choi,David Culler, Prabal Dutta, Cory Sharp, Gilman Tolle, MikeGrimmer, Bill Ferriera, and Ken Parker. Exscal: Elements ofan extreme scale wireless sensor network. In Proceedings ofthe 11th IEEE International Conference on Embedded andReal-Time Computing Systems and Applications(RTCSA),2005.

9. Taposh Banerjee, V. Kavitha, and V. Sharma. Energy efficientchange detection over a MAC using physical layer fusion,. InProc. IEEE International Conference on Acoustics, Speech, andSignal Processing (ICASSP), Las Vegas, USA, 2008. IEEE.

10. Luca Benini, Giuliano Castelli, Alberto Macii, and RiccardoScarsi. Battery-driven dynamic power management. IEEE Des.Test, 18(2):53–60, 2001.

11. A. B. Bhaskar, Girish Krishnan, N. Shamsudhin, and G.K.Ananthasuresh. Design, fabrication, and tetsing of a meso-scale accelerometer made of spring steel. Journal of theInstrumentation Society of India, 39(1):46–52, March 2009.

12. Abhijit Bhattacharya and Anurag Kumar. Delay constrainedoptimal relay placement for planned wireless sensor networks.In Proc. IWQoS 2010: The 18th IEEE International Workshopon Quality of Service. IEEE, June 2010.

13. Bernhard E. Boser, Isabelle M. Guyon, and Vladimir N. Vapnik.A training algorithm for optimal margin classifiers. In Proc.of the Fifth Annual Workshop on Computational LearningTheory, pages 144–152. ACM Press, 1992.

14. B. Bougard, F. Catthoor, D. C. Daly, A. Chandrakasan, andW. Dehaene. Energy efficiency of the ieee 802.15.4 standardin dense wireless microsensor networks: Modeling andimprovement perspectives. In Design Automation and Testin Eurpoe (DATE), March 2005.

15. Michael Buettner, Gary V. Yee, Eric Anderson, and RichardHan. X-mac: a short preamble mac protocol for duty-cycledwireless sensor networks. In SenSys ’06: Proceedings of the4th international conference on Embedded networked sensorsystems, pages 307–320, New York, NY, USA, 2006. ACM.

16. Costas Busch, Srikanth Surapaneni, and Srikanta Tirthapura.Analysis of link reversal routing algorithms for mobile ad hocnetworks. In SPAA ’03: Proceedings of the fifteenth annualACM symposium on Parallel algorithms and architectures,pages 210–219, New York, NY, USA, 2003. ACM.

17. Chih-Chung Chang and Chih-Jen Lin. Libsvm: a libraryfor support vector machines. Software available athttp://www.csie.ntu.edu.tw/∼cjlin/libsvm.

18. Elaine Cheong and Jie Liu. galsc: A language for event-drivenembedded systems. In DATE ’05: Proceedings of the conferenceon Design, Automation and Test in Europe, pages 1050–1055,Washington, DC, USA, 2005. IEEE Computer Society.

19. A. Dunkels, B. Gronvall, and T. Voigt. Contiki — a lightweightand flexible operating system for tiny networked sensors.In Local Computer Networks, 2004. 29th Annual IEEEInternational Conference on, pages 455–462, Nov. 2004.

20. A. Duwel and Barbour. Mems development at draperlaboratory. In Proceedings of the SEM Annual Conference,Charlotte, NC, June 2–4 2003.

21. Satyam Dwivedi. Low Power Receiver Architecture andAlgorrithms for Low Rate Wireless Personal Area Network. PhDThesis, IISc, 2010.

22. Virantha Ekanayake, IV Clinton Kelly, and Rajit Manohar. AnUltra Low-Power Processor for Sensor Networks. SIGARCHComput. Archit. News, 32(5):27–36, 2004.

23. A. Ekimov and Sabatier, J.M. Ultrasonic wave generation dueto human footsteps on the ground. Journal of the AcousticalSociety of America, 21(3):EL114–119, March 2007.

24. N. Furestenau, M. Schmidt, H. Horack, W. Goetze, andW. Schmidt. Extrinsic fabry-perot interferometer vibrationacoustic senso systems for airport ground traffic monitoring.In Proc. IEE on Optoelectronics, volume 144, pages 134–144,1997.

25. Eli M. Gafni, Student Member, Dimitri, P. Bertsekas, andSenior Member. Distributed algorithms for generating loop-free routes in networks with frequently changing topology.IEEE Transactions on Communications, 29:11–18, 1981.

26. David Gay, Philip Levis, Robert von Behren, Matt Welsh,Eric Brewer, and David Culler. The nesc language: A holisticapproach to networked embedded systems. In PLDI ’03:Proceedings of the ACM SIGPLAN 2003 conference onProgramming language design and implementation, pages1–11, New York, NY, USA, 2003. ACM.

27. Lin Gu, Dong Jia, Pascal Vicaire, Ting Yan, Liqian Luo, AjayTirumala, Qing Cao, Tian He, John A. Stankovic, TarekAbdelzaher, and Bruce H. Krogh. Lightweight detectionand classification for wireless sensor networks in realisticenvironments. In Proceedings of the 3rd internationalconference on Embedded networked sensor systems (SenSys),Nov. 2005.

28. A. Gupta and J. Kuri. Deterministic Schemes for KeyDistribution in Wireless Sensor Networks. In Proceedingsof the Third International Conference on COMmunicationSystem softWAre and middlewaRE (COMSWARE, Bangalore,India, January 2008.

29. A. Gupta, J. Kuri, and P. Nuggehalli. A New Scheme for



Establishing Pairwise Keys in a Wireless Sensor Network. InLecture Notes in Computer Science (LNCS), InternationalConference on Distributed Computing and Networking(ICDCN), December 2006.

30. A. Gupta, P. Nuggehalli, and J. Kuri. An Efficient Schemefor Establishing Pairwise Keys in a Wireless Sensor Network.In Proceedings of the First Workshop on Wireless Systems:Advanced Research and Development (WISARD), Bangalore,India, January 2007.

31. Tian He, Sudha Krishnamurthy, Liqian Luo, Ting Yan, Lin Gu,Radu Stoleru, Gang Zhou, Qing Cao, Pascal Vicaire, John A.Stankovic, and Tarek F. Abdelzaher. Vigilnet: An IntegratedSensor Network System for Energy-Efficient Surveillance.ACM Mobisys, 2004.

32. S. Hegde, Thejas, and N. Bhat. Universal capacitance sensor.In International conference on smart structures and Systems,pages 59–66, June 2008.

33. Wendi Rabiner Heinzelman, Anantha Chandrakasan, and HariBalakrishnan. Energy-efficient communication protocol forwireless microsensor networks. In Proceedings of 33rd AnnualHawaii International conference on System Scineces, 2000.

34. Mark Hempstead, Nikhil Tripathi, Patrick Mauro, Gu-Yeon Wei, and David Brooks. An Ultra Low Power SystemArchitecture for Sensor Network Applications. In Proceedingsof the 32nd annual international symposium on ComputerArchitecture, pages 208–219. IEEE Computer Society, 2005.

35. Jason Hill, Robert Szewczyk, Alec Woo, Seth Hollar, DavidCuller, and Kristofer Pister. System architecture directionsfor networked sensors. SIGPLAN Not., pages 93–104, 2000.

36. Hougen, D. F. et al. A miniature robotic system forreconnaissance and surveillance. In Proc. 2000 IEEEInternational Conference on Robotics and AutomationSan,Francisco, pages 501–507, April 2000.

37. Arunkumar Jayaprakasam and V. Sharma. Cooperative robustsequential detection algorithms for spectrum sensing incognitive radio. In Proc. International Conference on UltraModern Telecommunication (ICUMT), St. Petersburg, Russia,October 2009.

38. H. M. Yuichi Kaji and R. Matsumoto. Key PredistributionSchemes for Wireless Sensor Networks using Lines and Pointsover a Finite Geometry. In Proceedings of IEEE SECON,September 2006.

39. Brad Karp and H. T. Kung. Gpsr: greedy perimeter statelessrouting for wireless networks. In MobiCom ’00: Proceedings ofthe 6th annual international conference on Mobile computingand networking, pages 243–254, New York, NY, USA, 2000.ACM Press.

40. S. Khan, Thejas, N. Bhat, and G.K. Ananthasuresh. Designand characterization of a micromachined accelerometer withmechanical amplifier for intrusion detection. In 3rd ISSSNational Conference on Smart and Micro Systems, Kolkata,November 2009.

41. Shinji Kirihata, Katsuhiro Uchisawa, and Masao Yamaguchi.Low-profile dome-shaped multilens system. U.S. Patent6,051,836, Apr. 18 2000.

42. G. Krishnan. Displacement-amplifying compliant mechanismsfor sensor applications.

43. G. Krishnan and G. K. Ananthasuresh. A systematic methodfor the objective evaluation and selection of compliantdisplacement amplifying mechanisms for sensor applications.Journal of Mechanical Design, 130(Issue 10):102304:1–9, 2008.

44. G. Krishnan, C.U. Kshirasagar, and G.K. Ananthasuresh.Micromachined high-resolution accelerometers. Journal ofthe Indian Institute of Science: a Multidisciplinary ReviewsJournal, 87(3):333–361, 2007.

45. Walter Kuster and Hans J. Keller. Domed segmented lenssystem. U.S. Patent 4,930,864, Jun. 5 1990.

46. Zhao Lei, Zhang W.-H., Xu C.-N., Xu Y.-J., and Li X.-W.Energy-Aware System Design for Wireless Sensor Network.ACTA AUTOMATICA SINICA, 32:892–899, 2006.

47. Ben Leong, Barbara Liskov, and Robert Morris. Geographicrouting without planarization. In NSDI’06: Proceedingsof the 3rd conference on Networked Systems Design &Implementation, pages 25–25, Berkeley, CA, USA, 2006.USENIX Association.

48. Philip Levis, David Gay, Vlado Handziski, Jan-HeinrichHauer, Ben Greenstein, Martin Turon, Jonathan Hui, KevinKlues, Cory Sharp, Robert Szewczyk, Joseph Polastre, PhilipBuonadonna, Lama Nachman, Gilman Tolle, David Culler, andAdam Wolisz. T2: A second generation os for embedded sensornetworks. Technical report, Telecommunication NetworksGroup, Technische Universitat Berlin, November 2005.

49. H. Li and A. Lal. Radioisotape-Powered Cantilever for VacuumSensing with RF Transmission. In Proceedings of 12thInternational Conference on Transducers, Solid-State Sensors,Actuators and Microsystems, 2003.

50. H. Ling and T. Znati. End-to-end Pairwise Key Establishmentusing Multi-path in Wireless Sensor Networks. In Proceedingsof Globecom, December 2005.

51. J.C. Lotters, W. Olthuis, P.H. Veltink, and P. Bergveld. Asensitive differential capacitance to voltage converter for sensorapplications. IEEE Transactions on Instrumentation andMeasurement, 48(1), February 1999.

52. Gopal Pandurangan Maleq Khan and V.S. Anil Kumar.Distributed algorithms for constructing approximateminimum spanning trees in wireless sensor networks. InParallel and Distributed Systems, IEEE Transactions, volume 20,January 2009.

53. G.P. Mazarakis and J.N Avaritsiotis. A prototype sensornode for footstep detection. In Proceedings of the SecondEuropean Workshop on Wireless Sensor Networks. DOI:10.1109/EWSN.2005.1462036, January 2005.

54. Yajun Mei. Information bounds and quickest change detectionin decentralized decision systems. IEEE Transactions onInformation theory, 51(7):2669–2681, July 2005.

55. R. N. Miles, R. Robert, and R. R. Hoy. Mechanically coupledears for directional hearing in parasitoid fly ornia ochracea.Journal of the Acoustical Society of America, 98(6):3059–3070,1995.

56. K. Miller, A. Cowen, G. Hames, and B. Hardy. Soimumpsdesign handbook, memscap, revision 4.0., 2004.

57. MS3110 Data sheets. http://www.irvine-sensors.com/pdf/ms3110%20datasheet%20USE.pdf.

58. Raghuraman Mudumbai, Gwen Barriac, and UpamanyuMadhow. On the feasibility of distributed beam formingin wireless networks. IEEE Transactions on WirelessCommunications, 6(5):1754–1763, May 2007.

59. Leyla Nazhandali, Bo Zhai, Javin Olson, Anna Reeves, MichaelMinuth, Ryan Helfand, Sanjay Pant, Todd Austin, and DavidBlaauw. Energy Optimization of Subthreshold-Voltage SensorNetwork Processors. SIGARCH Comput. Archit. News,33(2):197–207, 2005.

60. I. V. Nikiforov. A generalized change detection problem. IEEETransactions on Information theory, 41(1):171–187, Jan 1995.

61. E. S. Page. Continuous inspection schemes. Biometrika,41(1/2):100–115, June 1954.

62. Panasonic Electric Works Corporation of America, New Jersey,USA. MP Motion Sensor (AMN 1,2,4) Data Sheet.

63. Joseph Polastre, Jason Hill, and David Culler. Versatilelow power media access for wireless sensor networks. InProceedings of the 2nd international conference on Embeddednetworked sensor systems, pages 95–107. ACM, 2004.

64. K. Premkumar and Anurag Kumar. Optimal sleep-wakescheduling for quickest intrusion detection using sensornetworks. In Proc. IEEE Infocom, Arizona, USA, Apr 2008.

65. K. Premkumar, Anurag Kumar, and Joy Kuri. Distributeddetection and localization of events in large ad hoc wirelesssensor networks. In Proc. Forty-Seventh Annual AllertonConference on Communication, Control, and Computing, 2009.

66. K. Premkumar, Anurag Kumar, and Venugopal V. Veeravalli.



Bayesian quickest transient change detection. In Proc. FifthInternational Workshop on Applied Probability (IWAP), 2010.

67. K. Premkumar, V. K. Prasanthi, and Anurag Kumar. Delayoptimal event detection on ad hoc wireless sensor networks.ACM Transactions on Sensor Networks. Under Second Review.

68. Adi Mallikarjuna V. Reddy, A.V.U. Phani Kumar, D. Janakiram,and G. Ashok Kumar. Wireless sensor network operatingsystems: a survey. International Journal of Sensor Networks,5(4):236–255(20), Aug 2009.

69. Curt Schurgers, Olivier Aberthorne, and Mani Srivastava.Modulation scaling for Energy Aware Communication Systems.In Proceedings of the international symposium on Low powerelectronics and design, pages 96–99. ACM, 2001.

70. A. N. Shiryaev. Optimal Stopping Rules. Springer, New York,1978.

71. Victor Shnayder, Mark Hempstead, Bor rong Chen,Geoff Werner Allen, and Matt Welsh. Simulating the PowerConsumption of Large-Scale Sensor Network Applications. InProceedings of the 2nd international conference on Embeddednetworked sensor systems, pages 188–200. ACM, 2004.

72. Roberto Solis, Vivek S. Borkar, and P. R. Kumar. A newdistributed time synchronization algorithm for multihopwireless networks. In Proceedings of the 45th IEEE Conferenceon Decision and Control, pages 2734–2739, San Diego,December 2006.

73. C. Srisathapornphat and Chien-Chung Shen. CoordinatedPower Conservation for Ad hoc Networks. In Proceedings of

IEEE International Conference on Communications, April2002.

74. A.G. Tartakovsky and V.V. Veeravalli. Quickest changedetection in distributed sensor systems. In Proc. SixthInternational Conference of Information Fusion, volume 2,pages 75–763, 2003.

75. Crossbow Technology. www.xbow.com.76. Vlasios Tsiatsis, Scott Zimbeck, and Mani Srivastava.

Architecture Strategies for Energy-Efficient Packet Forwardingin Wireless Sensor Networks. In Proceedings of theInternational symposium on Low power electronics and design,pages 92–95. ACM, 2001.

77. Ye Wei, J. Heidemann, and D. Estrin. Medium accesscontrol with coordinated adaptive sleeping for wireless sensornetworks. Networking, IEEE/ACM Transactions on, 12(3):493– 506, June 2004.

78. Wei Ye, J. Heidemann, and D. Estrin. An energy-efficient macprotocol for wireless sensor networks. pages 1567–1576, 2002.

79. Leena Zacharias and Rajesh Sundaresan. Decentralizedsequential change detection using physical layer fusion. IEEETransactions on Wireless Communications, 7(12):4999–5008,Dec. 2008.

80. Z. Zhang, Xuebin Gao, J. Biswas, and J. K. Wu. Moving targetsdetection and localization in passive infrared sensor networks.In The 10th International Conference on Information Fusion,July 2007.


reviews - ece.iisc.ac.inanurag/papers/anurag/kumar-etal10wsn-intrud… · indian institute of...

Documents