Evasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate
summary
Open-source Adversarial Robustness Toolbox (ART) demonstrates concrete attacks and defenses: attribute inference on the nursery dataset yields ~70% accuracy versus a 66% majority baseline; membership inference against a logistic regression gets 54% attack accuracy while the model is 91% accurate. Differential privacy via Diffprivlib drops model accuracy to ~56% and attack to ~51%, and tuning epsilon=25 finds a practical accuracy/privacy tradeoff. ART integrates with frameworks, supports evasion/poisoning/extraction/inference, and pairs with Armory and Counterfit for evaluations.
AI-written summary · about
chapters (7)
- 0:00Start
- 1:05Adversarial Threats against AI and ML
- 2:52Real-world Adversarial Exploits
- 5:03Extraction and Evasion against Email Protection System
- 6:32Adversarial Robustness Toolbox (ART)
- 12:51Attribute Inference Attack - Introduction
- 18:26Membership Inference Attack - Introduction
related talks
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
26mAttriGuard: A Practical Defense Against Attribute Inference Attacks via Adversarial Machine Learning
36mAdversarial Machine Learning: How Secure Is ML?
46mLessons Learned from Evaluating the Robustness of Defenses to Adversarial Examples
44mComputers are Stupid: Protecting "AI" from Itself
21mAdversarial Attacks
28mSecurity of Machine Learning
28mPrivacy preserving machine learning