Personal data de-identification for data science tasks
summary
Explains concrete tradeoffs: RSA ElGamal and Paillier style homomorphic encryption preserves additive or multiplicative ops so models can train on encrypted features but non linear activations require polynomial approximations, increasing cost. Describes tokenization as deterministic replacement that keeps identical inputs identical, generalization that keeps ML utility, and suppression that destroys it. Mentions Python package Faker, a rising tokenization market, real breaches at Marriott and British Airways, and latency and key management as the practical failure modes.
AI-written summary · about
chapters (10)
- 0:00Start
- 3:25Basic GDPR checklist (2)
- 10:30Sensitive data examples
- 11:50Personal data examples
- 12:45Personal data definitions
- 16:49Pseudonymization (tokenization) & anonymization
- 18:52Suppression & generalization
- 20:06Encryption & homomorphic encryption
- 20:57Python packages for data anonymization/tokenization
- 21:59Homomorphic encryption (1)
related talks
29mPersonal Data Pseudonymization Versus Anonymization In The Age Of AI & Big Data
21mThe Emperor's New Embeddings: Obfuscating ML Inputs Doesn't Provide Privacy
37mOpen-source Machine Learning on Encrypted Data
43mGuarding the ML Galaxy: Beyond Accuracy to Privacy and Security
33mPractical Data Privacy
52mMachine Learning over Encrypted Data with Fully Homomorphic Encryption
29mBalancing Thin Line Between Data Intelligence & Privacy
26mDoes using data mean giving up privacy?