Practical Data Privacy
summary
A practical claim: pure anonymity is mathematically impossible, so systems must trade information for privacy and choose techniques deliberately. A Netherlands fraud model used nationality as a decisive feature and produced biased outcomes. The 2020 US Census applies differential privacy (Tumult Analytics on Spark) to add noise and protect outliers. Google uses federated learning for keyboard prediction (2016–17) and private set intersection in Ads to match cohorts without centralizing raw data. Pseudonymization, hashing and tokenization remain weak links.
AI-written summary · about
chapters (16)
- 0:00Introduction to data privacy
- 1:32Fraud algorithm privacy concerns
- 4:06Defining data privacy
- 6:58Data governance foundations
- 10:10Bridging human privacy needs
- 12:18Data engineering privacy steps
- 15:08Pseudonymization techniques
- 17:39U.S. Census privacy failure
- 19:02Differential privacy explained
- 21:51Privacy information continuum
- 23:18Avoiding data collection
- 24:00Federated learning privacy
- 27:56Private set intersection
- 30:07Cryptography for privacy
- 31:10Privacy as a team sport
- 31:52Why privacy matters
related talks
29mPersonal Data Pseudonymization Versus Anonymization In The Age Of AI & Big Data
27mThe ethical algorithm
16mPrivacy by Design for JavaScript Data Systems
32mThe Dangerous Data Anonymization - Keynote
01h 17mTo be or not to be (anonymous)? Anonymyity in the age of big and open data.
29mPrivacy-preserving methods: Building secure projects
38mPrivacy: the Last Stand for Fair Algorithms
20mShare data without revealing personal information