隐私增强技术实践：差分隐私与去标识化部署.pdf

《隐私增强技术实践：差分隐私与去标识化部署.pdf》由会员分享，可在线阅读，更多相关《隐私增强技术实践：差分隐私与去标识化部署.pdf（35页珍藏版）》请在三个皮匠报告上搜索。

1、PETs JETs:Practical Differential Privacy&De-ID Deployment4:30 Wednesday April 3,2024Hal TriedmanSenior Privacy EngineerWikimedia FoundationWELCOME AND INTRODUCTIONSMiguel GuevaraProduct ManagerGoogleGerome MiklauCEO/FounderTumult LabsSarah CortesPrivacy EngineeringNetflixI.Welcome and Introductions

2、II.What is differential privacy,and how does it work?-Sarah Lewis CortesIII.Wikimedia Foundation Case Study-Hal TriedmanIV.Safely releasing earnings data using differential privacy-Gerome MiklauV.Example implementation-Miguel GuevaraVI.Questions and AnswersAGENDA OUTLINEWhat is differential privacy,

3、and how does it work?Sarah Lewis CortesWikimedia Foundation Case StudyHal TriedmanDP in practice:the Wikimedia Foundation(WMF)Transparency and open access to platform data are core values at WMFTransparency and open access to platform data are core values at WMFbut at the same time,so is user privac

4、y.WMFs Lean Data DietDefined by our Privacy Policy and Data Retention Guidelines:90 days until aggregation+deletionNo account neededNo first-party tracking cookies(images from Wikimedia Commons)DP geo-pageview releaseCommunity request:can WMF safely release data as possible about reading activity pa

5、rtitioned by both country and project?Recall:DP geo-pageview releaseWhat problem are we trying to solve?What does success look like?(broadly)-release as much data as possible about reading activity-partition by country,project,and page-release every day-Privacy protected at a user-day level-Data is

6、more plentiful and granular than baseline-Output is equitable,accurate,and trustworthy for data consumersList all possible keysSensitive data Public dataGroup by keyset+compute exact countsAdd DP noise to countsRemove low countsPublish the data!Compare exact and noisy counts to calculate errorLimit