AI in a Minefield: Learning from Poisoned Data
summary
Clear, concrete coverage of how training-data attacks work and real examples — fake review campaigns, spam-label poisoning against Gmail, and steganographic image poisons — followed by an actionable mitigation: a threshold learning design that only accepts signals seen from many distinct clients. The talk explains a streaming-friendly Boolean facts implementation (memory proportional to objects and properties) and its limits (rare-but-legitimate behaviors may be excluded), so viewers know what was built, why it works, and where it fails.
AI-written summary · about
chapters (6)
- 0:00AI risks and threats
- 10:08AI threats and data poisoning
- 20:55Securing web apps and APIs
- 31:40Threshold learning for anomaly detection
- 39:06Mitigating data poisoning in real-time
- 42:16Q&A
related talks
46mA Short Introduction to Hacking Artificial Intelligence
42mDefense Against the Dark Arts and Potions: AI Models Can Be Easily Poisoned
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
33mPoisoning Web-Scale Training Datasets is Practical
28mSecurity of Machine Learning
44mComputers are Stupid: Protecting "AI" from Itself
18mDefending the unknown using AI and cybersecurity
52mSleeping With One AI Open: An Introduction to Attacks Against...