Adversarial Attacks
summary
Covers concrete attack types and limits: gradient-based minimal perturbations (DeepFool-style, L1/L2 norms) that flip a panda to a gibbon with imperceptible noise, white-box vs black-box and data poisoning that corrupts training labels, publicized Google Cloud Vision evasion and antivirus evasion examples, and real harms like medical image tampering (altered X‑rays/blood results). Shows defenses—adversarial training, input filtering, robust preprocessing—are costly, brittle and part of an ongoing attacker–defender arms race.
AI-written summary · about
chapters (13)
- 0:00Adversarial attacks in AI
- 1:50Definition and impact
- 3:57Motivation and types
- 6:06How attacks work with examples
- 8:20Consequences and ethics
- 11:10Applications and vulnerabilities
- 13:18Case studies
- 14:48Future directions
- 16:10Conclusion
- 16:58Q&A: Perturbed inputs
- 18:22Q&A: Input filtering
- 19:05Q&A: Language model risks
- 20:11Q&A: Closed-system attacks
related talks
36mAdversarial Attacks on Deep Learning
38mUsing adversarial samples to robustify your Neural Network Models
25mTricking Neural Networks : Explore Adversarial Attacks
36mAdversarial Machine Learning: How Secure Is ML?
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
42mHow to attack AI systems (and how to defend them) !!!!
46mA Short Introduction to Hacking Artificial Intelligence
53mSecurity of Machine Learning Systems