Defense Against the Dark Arts and Potions: AI Models Can Be Easily Poisoned
summary
AI models suffer training‑time poisoning where injected samples either raise overall error (availability/error‑generic) or target specific classes (targeted/error‑specific). Attackers use optimal bilevel attacks that compute gradients w.r.t. training samples and expensive iterative optimization, or scale stealthy attacks with GAN‑generated stealthy poisons that look realistic yet subtly degrade accuracy and bypass distance‑based filters. Defenses include data filtering with a trusted validation set, stronger regularization/localization to harden training, and proactive techniques like unlearnable examples that add imperceptible noise to prevent learning.
AI-written summary · about
chapters (26)
- 0:00Intro
- 0:38Bad News
- 1:13Who am I
- 1:55Applications of AI
- 2:44Special Example
- 6:09Data Poisoning Examples
- 7:10What is a Person Attack
- 8:17Why Should You Worry
- 9:12Data Poisoning Attacks
- 10:49Microsoft Survey
- 11:17Data Posting Attacks
- 12:40Evaluating Data Posting Attacks
- 13:26Data Pipeline Process
- 14:35Optimal Attacks
- 15:42Level Attention Problems
- 18:44Optimal Person Attacks
- 18:56The Gans
- 20:31The Classifier
- 25:21Data Data Filtering
- 26:09Data Validation
- 27:28Binary Classification
- 29:24unlearnable examples
- 30:31learnable examples
- 31:30takeaways
- 32:34Professor Jose
- 33:02Reference
related talks
46mAI in a Minefield: Learning from Poisoned Data
44mComputers are Stupid: Protecting "AI" from Itself
45mSecuring AI: There Is No Try, Only Do!
31mSecurity Considerations for Services Using AI Models
46mA Short Introduction to Hacking Artificial Intelligence
20mMachine Learning Attacks – A New Era of Security Epidemic
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence