Sleeping With One AI Open: An Introduction to Attacks Against...
summary
A taxonomy of practical ML attacks with live demos and tool names: model theft via query-based distillation (Knockoff Nets reproduced a model for ~$30), decision-boundary evasion with adversarial examples (one-pixel attacks and stickered stop signs breaking vision), and model hijacking that embeds malware by altering serialized weights. Demonstrated tools include Counterfeit, a HopSkipJump-based HopSkipJump evasion run against a credit-fraud model, and a working pickle steganography pipeline that embeds and executes ransomware when a Torch model is loaded.
AI-written summary · about
chapters (13)
- 0:00Introduction to AI risks
- 3:10AI vs machine learning basics
- 8:06Who targets AI systems
- 11:17Adversarial ML tools overview
- 13:47Attacker goals and timing
- 17:01Poisoning attacks explained
- 23:21Inference attacks deep dive
- 31:27Adversarial example demo
- 34:28Model hijacking attacks
- 40:02Model serialization risks
- 45:22Malware in ML models demo
- 49:39Mitigation strategies summary
- 50:43Closing remarks
related talks
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
46mA Short Introduction to Hacking Artificial Intelligence
37mKeynote - Hacking AI How to Survive the AI Uprising
53mSecurity of Machine Learning Systems
30mThe Simple, Yet Lethal, Anatomy of an AI Attack
44mHacking AI - How to Survive the AI Uprising
28mSecurity of Machine Learning
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications