A Short Introduction to Hacking Artificial Intelligence
summary
Describes practical adversarial attacks on machine learning: small physical stickers can fool autonomous vehicle sign classifiers, poisoning training data with glasses can implant backdoors into face-authentication models, and targeted poisoning with dozens of examples can create Berkeley-style backdoors. Presents scaling evidence (ImageNet error drop and a spam-filter study where ~10% text perturbation flips labels) and recommends defenses: data hygiene, robust training, and hardening ML libraries.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 10:06What is AI
- 25:56Machine learning basics
- 31:20AI attack methods
- 40:52AI system security
related talks
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
36mAdversarial Machine Learning: How Secure Is ML?
52mSleeping With One AI Open: An Introduction to Attacks Against...
28mSecurity of Machine Learning
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
36mAdversarial Attacks on Deep Learning
44mComputers are Stupid: Protecting "AI" from Itself
53mSecurity of Machine Learning Systems