databin data’drivensecure- businessintelligence€¦ · majorchallenges...

19
DataBIN DataDriven Secure Business Intelligence Devdatt Dubhashi David Sands

Upload: others

Post on 26-Jul-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

DataBINData-­‐Driven  Secure  Business  Intelligence

Devdatt DubhashiDavid  Sands

Page 2: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-
Page 3: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Major  Challenges

• How  do  we  automatically  extract  meaningful  info  from  unstructured  text,  images,  video  …

• How  do  we  structure  the  information  for  better  data  analytics?

• How  do  we  scale  to  very  Big  Data?• How  do  we  ensure  privacy  when  mining  info?

Page 4: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

• Graph  kernels  for  network  structured  data,  ICML  2014,  NIPS  2015,  KDD  2015,  CIKM  Weighted  Theta  Functions,  NIPS 2015  

• Large  scale  optimization:  clustering, domain  adaptation,  ICML  2017…

• Explanatory  AI/ML:  Causal  and  Counterfactual  inference,  ICML  2016,  ICML  2017

• Explanatory  AI/ML:  Disentangled  representations  in  deep  nets.

• Deep  Learning  for  NLP:  char  based  RNNs.

• Differential  Privacy:    JMLR 2017,  AAAI  2017

1Disciplinary  research  published  at  top-­‐tier  conferences  

Page 5: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Demonstrators  implemented  and  integrated  into  the  tools  of  our  industrial  partners  

2

Page 6: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Dissemination“AI  is  the  New  Electricity”

3

Swedish  Symposium  Deep  Learning  2018

Page 7: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Competence  Intelligence

Innovation

Page 8: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Privacy  in  the  Age  of  Big  Data

Page 9: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-
Page 10: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

“Two  recent  surveys  reveal  that  consumers’  concerns  about  online  privacy  are  at  an  all-­‐time  high.” June  2014

“Big  data  might  be  big  business,  but  overzealous  data  mining  can  seriously  destroy  your  brand…”                        

Nov  2013

Page 11: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Research  on  Privacy  in  Data-­‐Intensive  Systems   Differential  Privacy

Location  Privacy

Social  Network  Privacy

Page 12: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

A  Flavour  of  Differential  Privacy

A  personal  question…

Page 13: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

13

Page 14: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

14

Answer  YES

Page 15: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

15

Answer  YES

Answer  NO

Page 16: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

16

Answer  YES

Answer  NO

Answer  TRUTHFULLY

Page 17: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Differential  Privacy

Emerging  mathematical  definition  of  privacy

Essence: the  participation  of  any  one  individual  won’t  change  the  result  of  the  survey  in  a  noticeable  way

Consequence:  a  robust  definition  with  good  properties

17

Page 18: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

Results  in  the  DataBIN Project

• Programming  framework  that  achieve  privacy  by  construction– no  need  to  trust  the  programmer

• A  Framework  for  Local  Differential  Privacy– no  need  to  trust  the  analyst

• Machine  Learning  with  Differential  Privacy

Page 19: DataBIN Data’DrivenSecure- BusinessIntelligence€¦ · MajorChallenges •How-do-we-automatically-extract-meaningful-info-from-unstructured-text,images,video-… •How-do-we-structure-the-informationfor-

DataBIN PhDs

Olof MogrenDeep  Learning  NLP

Hamid  EbadiDifferential  Privacy Raul  Pardo  (INRIA  Lyon)

Privacy  in  Social  Networks

Fredrik  Johansson  (MIT)Machine  Learning,  Causal  Inference

See  Posters