Adversarial Examples
summary
Adversarial robustness remains unsolved empirically and theoretically: on CIFAR‑10 with epsilon=8 the best reported robust accuracy is ~57% and on ImageNet a top result is 43%. Key claims: norm choice matters (L∞ vs L2), adversarial loss replaces loss by sup over perturbations, and standard VC/Rademacher bounds pick up dimension terms so you may need ~√d more samples; some constructions force computational hardness via pseudo‑random generators. Practical tools that work or certify include adversarial training and randomized smoothing for provable L2 guarantees.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:47Objective
- 2:29Surprising Fact
- 7:23The Talk
- 11:40SLT Framework
- 15:10XPRIZE
- 25:03Loss classes
- 31:15VC dimension
- 34:37Complexity
related talks
47m1Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers
49mA New Perspective on Adversarial Perturbations
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
46mLessons Learned from Evaluating the Robustness of Defenses to Adversarial Examples
09mLearning Under Requirements
12mDistribution-Free Robustness
59mStudying Failure Modes of Deep Learning and Ways To Fix Them
34mDistributional Robustness, Learning, and Empirical Likelihood