Adversarial Training

### Adversarial Attacks Against Computer Vision Systems At 00:00, Wenqing Cao and Yaohua Guo present their winning competition methodology for generating targeted adversarial images against deep neural networks at the AI Village. They explain how minor input perturbations deceive state-of-the-art vision models.

### Gradient-Based Perturbation Generation At 02:20, the speakers break down gradient-based attack mechanics, including Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). By computing gradients of model loss functions with respect to input pixels, attackers craft high-frequency, imperceptible image noise that forces targeted misclassifications.

### Black-Box Attack Transferability Across Architectures At 04:56, Cao and Guo address black-box attack scenarios where model architectures and weights are hidden from the attacker. They explain how training diverse ensembles of surrogate models generates perturbations that generalize and transfer successfully across unobserved target classifiers.

### Optimization Under Budget Constraints and Submission Limits At 06:55 and 11:22, the presenters share practical optimization strategies used during competition rounds under strict rate limits. They demonstrate how fine-tuning step sizes and gradient momentum produces robust adversarial samples with minimal query expenditures.

(AI summary — Gemini Flash)