experimental results of data leakage detection

11
EXPERIMENTAL RESULTS OF DATA LEAKAGE DETECTION In this section we provide experimental results of our project. The main thing in experimentation is to estimate how much different allocation techniques, that we have implemented, are efficient in identifying the correct ‘guilty’ agent. Experimental Setup: Before going to undertake experimentation let us consider following experimental arrangement: 1. There is one distributor 2. There are four agents U 1, U 2 , U 3 , U 4 3. There is one customer 4. Distributor has total 30 profiles in his database ( i.e. |T| = 30 ) Experiments: So to begin with, we have to consider different cases for both ‘Sample Request’ and ‘Explicit Request’. We show results in terms of graph. A) Sample Request Here we have to test implementation of both ‘s- random’ and ‘s- overlap’ algorithms. Case I: When M > |T|, where M = ∑ i=1,..,n m i ( i.e. total of number of profiles requested by all agents is more than available profiles) Let us consider following scenario: Agent (U i ) Number of profiles requested from distributor (m i ) Number of profiles given to customer U 1 8 5 U 2 7 - U 3 10 10 U 4 10 - Table 7.1. Scenario 1 For Sample Request Here, M = 35. i.e. M > |T|

Upload: parvezahmed-kazi

Post on 28-Apr-2015

113 views

Category:

Documents


3 download

DESCRIPTION

Experimental Results of Data Leakage Detection

TRANSCRIPT

Page 1: Experimental Results of Data Leakage Detection

EXPERIMENTAL RESULTS OF DATA LEAKAGE

DETECTION

In this section we provide experimental results of our project. The main thing in

experimentation is to estimate how much different allocation techniques, that we have

implemented, are efficient in identifying the correct ‘guilty’ agent.

Experimental Setup:

Before going to undertake experimentation let us consider following experimental

arrangement:

1. There is one distributor

2. There are four agents U1, U2, U3, U4

3. There is one customer

4. Distributor has total 30 profiles in his database ( i.e. |T| = 30 )

Experiments:

So to begin with, we have to consider different cases for both ‘Sample Request’ and ‘Explicit

Request’. We show results in terms of graph.

A) Sample Request

Here we have to test implementation of both ‘s-random’ and ‘s-overlap’ algorithms.

Case I: When M > |T|, where M = ∑i=1,..,n mi

( i.e. total of number of profiles requested by all agents is more than available profiles)

Let us consider following scenario:

Agent

(Ui)

Number of profiles requested

from distributor (mi)

Number of profiles given to

customer

U1 8 5

U2 7 -

U3 10 10

U4 10 -

Table 7.1. Scenario 1 For Sample Request

Here, M = 35. i.e. M > |T|

Page 2: Experimental Results of Data Leakage Detection

After executing the ‘Guilt Model’ on above scenario we get following results.

1) Result for s-random:

Graph 1.1.1: Guessing Probability (p) = 0.3

Graph 1.1.2: For all values (0 to 1) of Guessing Probability (p)

Page 3: Experimental Results of Data Leakage Detection

2) Result for s-overlap:

Graph 1.2.1: Guessing Probability (p) = 0.3

Graph 1. 2.2: For all values (0 to 1) of Guessing Probability (p)

Page 4: Experimental Results of Data Leakage Detection

By considering Scenario No.1( shown in Table 1.1) and the results obtained for both of

allocation strategies, s-random and s-overlap, it becomes clear that agents U3 and U1 are ‘more

guilty’ than other two agents U2 and U4. To be more specific, agent U3 is guiltier than others.

Case II: When M < |T|, where M = ∑i=1,..,n mi

( i.e. total of number of profiles requested by all agents is less than available profiles)

Let us consider following scenario:

Agent

(Ui)

Number of profiles requested

from distributor (mi)

Number of profiles

given to customer

U1 8 8

U2 7 -

U3 8 5

U4 6 -

Table 7.2. Scenario 2 For Sample Request

Here, M = 29. i.e. M < |T|

After executing the ‘Guilt Model’ on above scenario we get following results, we show results in

terms of graph.

1) Result for s-random:

Graph 2.1.1: Guessing Probability (p) = 0.3

Page 5: Experimental Results of Data Leakage Detection

Graph 2.1.2: For all values (0 to 1) of Guessing Probability (p)

2) Result for s-overlap:

Graph 2.2.1: Guessing Probability (p) = 0.3

Page 6: Experimental Results of Data Leakage Detection

Graph 2.2.2: For all values (0 to 1) of Guessing Probability (p)

By considering Scenario No.2( shown in Table 2.1) and the results obtained for both of

allocation strategies, s-random and s-overlap, it becomes clear that agents U1 and U3 are ‘more

guilty’ than other two agents U2 and U4. To be more specific, agent U1 is guiltier than others.

From several experimental results for above two cases, it also becomes clear that, the ‘s-

overlap’ allocation strategy gives more accurate results than ‘s-random’ strategy.

Page 7: Experimental Results of Data Leakage Detection

B) Explicit Request

In this, we have to consider the case in which distributor doesn’t use the fake object(s) (i.e. fake

profile(s)); in addition to case where it makes use of them. And also we provide results for both

implementations of algorithms ‘e-random’ and ‘e-optimal’.

Let us consider following scenario:

Agent (Ui) Condition 1 Condition 2 Number of profiles

requested from

distributor (mi)

Number of

profiles given

to customer

U1 Kolhapur ME 8 -

U2 Kolhapur BCS 8 8

U3 Kolhapur MCom 8 -

U4 Kolhapur MTech 8 -

Table 7.3 Scenario For Explicit Request

Case I: Distributor doesn’t use fake object

Graph: Guessing Probability (p) = 0.3

Graph: For all values (0 to 1) of Guessing Probability (p)

Page 8: Experimental Results of Data Leakage Detection

The result shows that, it is very difficult to identify the correct leaker if the distributor do

not makes use of fake object(s). This is because all agents receive same set of objects.

Now, let’s see the results when the distributor adds some fake objects to set of objects

received by individual agents.

Case II: Distributor uses fake objects

1) Result for e-random

Graph: Guessing Probability (p) = 0.3

Page 9: Experimental Results of Data Leakage Detection

Graph: For all values (0 to 1) of Guessing Probability (p)

In e-random, as we have mentioned earlier, while allocating fake object, agent is chosen

randomly from the set of agents that can receive fake objects. Looking at above two results it is

quite clear that we get improved results as compared to the case-I. But, still this result fails to

clearly distinguish between ‘guilty agent’ and ‘innocent agent’, that is here we get the same

‘guilt probability’ values for agents U1, U3 and U2, U4. This happened because of random

selection of agent.

2) Result for e-optimal

Here, instead of selecting randomly, agent is selected (for fake object allocation) in such a way

that will improve our chances to identify him/her if it leaks data.

Graph: Guessing Probability (p) = 0.3

Page 10: Experimental Results of Data Leakage Detection

Graph: For all values (0 to 1) of Guessing Probability (p)

From result, we are sure that agent U2 is the leaker.

Youtube Video : Click Here If anyone interested to buy the project, then contact on : [email protected]

Page 11: Experimental Results of Data Leakage Detection