Publishing Wikipedia Project Usage Data with Strong Privacy Protections and w/o Tracking
summary
Wikimedia Foundation pilots releasing page-view counts by project and country with differential privacy on a 1.5–2 TB/day dataset, working with Tumult Labs. Current user keying is an MD5 actor signature of IP+user agent, which breaks when IPs change and collapses distinct users on shared networks. The prototype replaces that with anonymous client-side filtering using a daily rotated global salt and a three-character hex fingerprint (4,096 values), including up to 10 page-views per device; collision estimates ~1.1% and open questions about guarantees and deployment remain.
AI-written summary · about
chapters (7)
- 0:00Intro
- 1:01The Wikimedia Foundation (WMF)
- 3:07Existing privacy methods
- 3:46WMF's Lean Data Diet
- 5:25Country-project-page release
- 8:57Fundamental tension
- 12:02Anonymous client-side filtering
related talks
18mPrivacy-Preserving Web Search
01h 04mAcademic sessions privacy and surveillance
49mWeb Side Story: A Privacy Odyssey through Tracking and Fingerprinting
36mCutting Gordian's "End-User Focus" vs. "Privacy" Knot
20mShare data without revealing personal information
37mWikimedia Content API: A Cassandra Use-case
32mThe Dangerous Data Anonymization - Keynote
15mHow We Can Save Anonymization