Do AI-models gossip?
summary
A practical toolkit called LeakPro runs attack-based risk assessments (membership inference, reconstruction, linking) against models exposed via API, federated updates, or synthetic outputs. A hospital length-of-stay model gains ~4% accuracy with a larger neural net while membership tests give an attacker ~8× advantage versus a logistic model. A PII-removal classifier trained on ~800 court documents allows an attacker to correctly flag 37 training records. The workflow measures probability a record was in training, attempts reconstruction, and repeats after adding privacy-enhancing tech.
AI-written summary · about
chapters (12)
- 0:00Introduction to AI gossip
- 1:12Types of AI model attacks
- 5:25Data leakage risks explained
- 7:34Privacy formal vs experimental
- 11:08LeakPro privacy tool overview
- 15:45Membership inference attacks
- 18:37Federated learning privacy
- 21:33Synthetic data linkage risks
- 22:16Healthcare use case study
- 26:13Leakage in API access models
- 30:29Court document PII removal
- 34:02Privacy utility tradeoff summary
related talks
28mPrivacy preserving machine learning
13mHow Cloud AI Leaks What It Learns
17mAI Safety in the Generative AI Era
43mGuarding the ML Galaxy: Beyond Accuracy to Privacy and Security
45mCombating Generative AI's Privacy Abuses
52mSleeping With One AI Open: An Introduction to Attacks Against...
20mMachine Learning Attacks – A New Era of Security Epidemic
28mStealing Machine Learning Models via Prediction APIs