Navigating the Security and Privacy Landscape of Modern AI
summary
Modern AI breaks three ML assumptions—future equals past, training data is representative, and data generation is independent of the model—causing predictable failures and security gaps; a psychology study finds 40% of people draw unusable bicycles, illustrating high accuracy without understanding. Concrete failures include intrusion detectors learning environment-specific shortcuts, a vulnerability scanner flagging code due to variable names not buffer size, physical sticker attacks that reliably fool classifiers, and LLM privilege hierarchy defenses that remain probabilistic. Mitigations: explainable AI, out‑of‑distribution testing, differential privacy (theoretical but hard in practice), runtime detection and red teaming.
AI-written summary · about
chapters (6)
- 0:00AI security and privacy intro
- 13:43Why AI fails without intent
- 21:54AI under attacks
- 35:02Privacy risks in AI
- 49:52LLM security pitfalls
- 1:02:15Future of AI security
related talks
45mSecuring AI: There Is No Try, Only Do!
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
28mSecurity of Machine Learning
49mAn AI with an Agenda
46mA Short Introduction to Hacking Artificial Intelligence
42mForward Focus – The Impact of Artificial Intelligence
45mWhat Security Measures Are Necessary to Trust in AI?
44mComputers are Stupid: Protecting "AI" from Itself