AttriGuard: A Practical Defense Against Attribute Inference Attacks via Adversarial Machine Learning
summary
AttriGuard defends against attribute inference attacks that use ML classifiers to predict private user attributes (like location or political view) from public data (ratings, likes). It adapts adversarial-evasion techniques to compute minimal perturbations for each possible attribute value (Phase I), then samples a value from an optimized probability distribution and applies the corresponding perturbation (Phase II) by solving a constrained convex optimization; the method is tractable and evaluated on a real dataset against prior defenses.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:31OUTLINE
- 2:05Attribute Inference Attacks are Pervasive
- 2:58Existing Defenses
- 4:22Our Defense: AttriGuard
- 4:47Threat Model
- 6:17Challenges
related talks
40mEvasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate
36mAdversarial Machine Learning: How Secure Is ML?
43mGuarding the ML Galaxy: Beyond Accuracy to Privacy and Security
24mYou Are Who You Know and How You Behave: Attribute Inference Attacks...
44mComputers are Stupid: Protecting "AI" from Itself
28mSecurity of Machine Learning
46mA Short Introduction to Hacking Artificial Intelligence
21mAdversarial Attacks