Is your model private?
summary
Neural nets memorize real sensitive data so membership inference and extraction attacks can reveal training examples (research even recovered secrets from GPT‑2). Protection uses differential privacy applied during training via DPSGD: per‑example gradient clipping plus Gaussian noise, with privacy budget ε (papers suggest roughly 1–10 as a practical range) and a clear utility/privacy tradeoff. Production tools include Opacus and TensorFlow Privacy and real deployments include Google Gboard and Apple telemetry.
AI-written summary · about
chapters (8)
- 0:00Introduction to privacy risks
- 1:27Membership inference attacks
- 4:56Unintended memorization risks
- 6:01Differential privacy basics
- 12:30Differential privacy for neural networks
- 17:31Libraries for implementation
- 19:41Balancing privacy and utility
- 21:07Q&A session
related talks
28mPrivacy preserving machine learning
01h 07mLarge-Scale Private Learning, Part II
31mScaling privacy embedded distributed ML systems
01h 17mTutorial: Privacy, Part I
33mPrivacy of Decentralized Machine Learning
19mPractical privacy-preserving machine learning in Python
35mDifferential Privacy Made Practical
30mPrivate Generative Models: Balancing Privacy and Utility in AI