Guarding the ML Galaxy: Beyond Accuracy to Privacy and Security
summary
A run-through of real incidents and concrete defenses: the Cambridge Analytica leak of 87 million users and a $550M Facebook settlement motivate coverage of attacks like evasion, poisoning and membership inference; the defenders shown include audit tools ML Privacy Meter and ML Doctor, differential privacy implemented via gradient clipping and epsilon-noise, federated learning that exchanges parameter updates instead of raw data, synthetic and condensed datasets, and encryption (homomorphic/pseudo-anonymization); tradeoffs and failure modes (memorization, communication leakage, degraded accuracy) are quantified.
AI-written summary · about
chapters (11)
- 0:00ML privacy and security intro
- 2:29Cambridge Analytica scandal
- 8:06Privacy attacks in ML
- 10:54Defining ML privacy
- 13:44Auditing ML privacy risks
- 17:41Differential privacy
- 19:54Federated learning
- 23:05Synthetic data for privacy
- 31:38Privacy regulations
- 37:58Innovation vs privacy
- 41:20Q&A
related talks
31mScaling privacy embedded distributed ML systems
28mPrivacy preserving machine learning
28mSecurity of Machine Learning
46mGuarding the LLM Galaxy: Security, Privacy, and Guardrails in the...
01h 01mData Security in Machine Learning
44mPrivacy by Design Machine Learning
43mPanel: Privacy Design Patterns for AI Systems: Threats and Protections
08mLLMs Without Leaks: Keep Your Code, Data & IP