Privacy preserving machine learning
summary
Models leak real examples: GPT‑2 reproduces phone numbers, reconstruction attacks recover identifiable photos, and membership inference reveals presence in sensitive datasets. An open source tool Privacy Evaluator (a fork of IBM ART) implements black‑box and white‑box attacks, a per‑point risk score and a membership‑advantage metric for vulnerability testing. Defenses include differentially private SGD (TensorFlow Privacy / Opacus) and compression via pruning reduces membership leakage while quantization does not.
AI-written summary · about
chapters (6)
- 0:00Start
- 8:22Why important?
- 12:55Solution: Vulnerability testing
- 18:41Defences: Differential privacy
- 24:11Compression
- 24:59Practical challenges
related talks
31mIs your model private?
43mGuarding the ML Galaxy: Beyond Accuracy to Privacy and Security
01h 17mTutorial: Privacy, Part I
35mDo AI-models gossip?
28mSecurity of Machine Learning
21m"SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving
30mPrivate Generative Models: Balancing Privacy and Utility in AI
01h 28mPPML: Machine Learning on Data you cannot see