Reproduce “Adversarial Machine Learning at Scale” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexey Kurakin, Ian Goodfellow, Samy Bengio
Researchers at Google Brain and OpenAI scaled adversarial training to ImageNet-sized Inception v3 models, enhancing their robustness against one-step adversarial attacks. The study identified a "label leaking" effect in evaluation and characterized the transferability properties of various adversarial example generation methods.
Key insights
- Adversarial training at scale acts primarily as a specialized regularizer for adversarial robustness, typically incurring a slight reduction in clean image accuracy.
- A "label leaking" effect was discovered in FGSM-based evaluations, where using the true label in adversarial example construction inadvertently boosts the adversarially trained model's accuracy on those specific examples.
- Adversarial robustness for adversarially trained models consistently increases with model capacity, while non-adversarially trained models show an optimal capacity for robustness.
- Single-step adversarial examples are significantly more transferable across different models than iterative adversarial examples, which are more potent but less generalizable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.