EEXCESS or the challenge of privacy-preserving quality recommendations Benjamin Habegger, Nadia Bennani, Elöd Egyed-Szigmond, Omar Hassan Lyon University, CNRS, INSA-Lyon, LIRIS, UMR5205
EEXCESS – Project Enhancing Europe’s eXchange in Cultural Educational and Scientifjc reSources. Started: February 2013 Duration: 4 Years Budget: 7.05 Million EUR http://eexcess.eu/
EEXCESS – Partners
EEXCESS – Main problem Making existing qualilty content visible “Popular” long-tail content Long-tail content ≠ Quality content 4
EEXCESS - Objectives Content enrichment Personalized recommendation Privacy preservation 5
EEXCESS – Global challenges • Federated recommendation • Integration of multiple document sources • Mining for user profjles • Adapting user interfaces to context • Preserving user privacy 6
EEXCESS – Tradeofg ? Privacy Quality • Unlimited disclosure – User privacy is clearly at stake. – Does it improve quality ? How much ? 7
EEXCESS – Tradeofg ? Privacy Quality • Limited disclosure – Limit recommendation quality ? 8
EEXCESS – Simplifjed architecture Client Application Usage EEXCESS Privacy EEXCESS Mining Proxy EEXCESS Federated Recommender Mendeley Econbiz Search Search 9
Usage mining → build detailed user profjles
EEXCESS – User context • Social context • Activity context – Browsing history – Friends – Ongoing tasks – Neighbors – Reading history – Co-workers • Environmental context – Relatives – Temperature, Humidity, • Spatio-temporal context Light – Things, Services – Time • Personal context – Location – Weight, Pulse, Blood – Direction of movement pressure, Mood 11
EEXCESS – User profjle • Demographic information • Knowledge, background, skills – Knoweldge within a domain – Age (e.g. acquired by a student) – Gender – Professional expertise and – Relationship status skills – Address • Goals – ... – Short term goal (current task) • Interests – Long term goal – Professional interests • Behavior – Personal interests – Repetitive behaviors – Interest in commercial – History of user actions products 12
Recommendation → fjnd recommendations adapted to the user profjles
EEXCESS – Recommendation • Content-based • Collaborative fjltering • Demographic fjltering • ... 14
Privacy preserving recommendation → compromise between private information disclosure and quality
EEXCESS – Privacy questions • What profjle information is useful ? – For usage mining ? – For quality recommendations ? • How much and how detailed should this information be disclosed ? – To preserve privacy ? – To ensure recommendation quality ? • What control and feedback can we provide ? – To „measure“ the impacts of disclosure 16
EEXCESS – Privacy wish-list • Guarantees – Deterministic ? ● Guarantee that a particular piece of information doesn't leak – Risk of disclosure (including inference) ● Measuring risks and impacts of privacy breaches • Flexibility – User-dependant policy ● Alice and Bob may have difgerent requirements – Context-dependant policy ● Alice may have difgerent requirements at home and at work 17
EEXCESS – Privacy techniques • Anonymization – K-anonymity [Sweeney2002] • Difgerential Privacy [Dwork2006] • Hiding in the crowd • Distributed recommendation • ... 18
EEXCESS – Privacy challenges • Data mining, Big data analytics • External knowledge – De-anonymization [Narayanan2008] – Breaches w.o. participation [Dwork2006] 19
EEXCESS – Privacy challenges Privacy breach w.o. participation [Dwork2006] What is the average age ? ... D 38 36 Paul is 2 years less than Paul the average person in D 20
Ongoing work
EEXCESS – Current focus • Setup a test-framework – User interface – API's – Evaluation tools • Involve the user – Transparency ● What is going on ? ● What does the system know ? – Control ● Let users defjne their own policy – Feedback ● Show the impacts of user's policy 22
EEXCESS – Simplifjed architecture Client Application Usage EEXCESS Privacy EEXCESS Mining Proxy EEXCESS Federated Recommender Mendeley Econbiz Search Search 23
Privacy Plugin – Recommendations 24
Privacy Plugin – Profjle edition 25
Privacy Plugin – Profjle data collection • Mendely auth + profjle import 26
Privacy Plugin ● Recommendations ● Oauth interactions with ● Privacy settings Mendeley Client ● Profjle editing Application ● Privacy sandbox Usage Privacy EEXCES EEXCES Mining Proxy S S ● Receives EEXCES Federated Recommender recommendations S ● User profjle edits ● User browsing activity Mendeley Econbiz ● Recommendation requests Search Search 27
Privacy Plugin - Transparency 28
Privacy Plugin – Trancparency 29
Privacy Plugin – Control & Feedback 30
Privacy Plugin – Control & Feedback 31
Privacy Proxy - Protection ● User profjle edits ● User browsing activity Client Application ● Recommendation requests ● Relays recommendations Usage Privacy EEXCES EEXCES Mining Proxy S S ● Stores profjles + browsing activity ● Receives ● Applies privacy settings EEXCES Federated recommendations Recommender S ● Relays policy-respectful recommendation requests Mendeley Econbiz Search Search 32
Basic Federated Recommender Client ● Receives ● Returns interleaved Application recommendation requests recommendations Usage Privacy EEXCES EEXCES Mining Proxy S S ● Receives recommendation request ● User static profjle EEXCES Federated ● User recent activity Recommender S ● Searches Econbiz ● Maps “user context” into a + Mendeley weighted term-based query Mendeley Econbiz Search Search 33
EEXCESS – Summary • Quality – Recommendations adapted to the user • Preserving privacy – Guaranteeing respect of privacy policy • Scalability – Ensuring the whole system scales • Quantifjable measures 34
EEXCESS – Future work • Privacy-preserving search • Measuring – recommendation quality – privacy impacts of disclosure • Impacts of trustworthiness in peers 35
Contact benjamin.habegger@insa-lyon.fr http://www.linkedin.com/in/benjaminhabegger @b_habegger
Recommend
More recommend
