A New Perspective on Adversarial Perturbations
summary
Claim: adversarial examples come from non-robust features in the data that reliably predict labels but flip under tiny perturbations. Specific experiment: training on a dataset relabeled by adversarial perturbations still yields ~78% accuracy on the original task, showing those perturbations carry predictive signal. Practical effects: robust training and randomized smoothing suppress non-robust features, producing human-aligned saliency maps and faithful feature visualizations but raising training cost, reducing standard accuracy, and requiring more data; transferability follows because these features are dataset-level.
AI-written summary · about
chapters (18)
- 0:00Why do we love deep learning?
- 3:02Key Problem: Adversarial Perturbations
- 4:51ML via Adversarial Robustness Lens
- 10:17Human Perspective
- 10:39ML Perspective
- 12:53A Simple Experiment
- 17:06The Robust Features Model
- 18:58The Simple Experiment: A Second Look
- 24:04Human vs ML Model Priors
- 25:35New capability: Robustification
- 27:37A Natural Consequence: Transferability
- 29:37The Role of Robust Training
- 30:25New Take on Randomized Smoothing
- 30:52Robustness and Data Efficiency
- 33:28A Simple Theoretical Setting: Max Likelihood Gaussian Classification
- 38:34Robustness + Perception Alignment
- 44:14Robustness + CV Applications
- 45:03Adversarial examples arise from non-robust features in the data
related talks
01h 24mAdversarial Examples
36mAdversarial Attacks on Deep Learning
21mAdversarial Attacks
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
36mAdversarial Machine Learning: How Secure Is ML?
44mAdversarial attacks on deep learning models
42mBeyond Adversarial Approaches on Distributional Robustness
38mUsing adversarial samples to robustify your Neural Network Models