Reproduce “Strength-Adaptive Adversarial Training” on OpenResearch

About this paper

Chaojian Yu, Dawei Zhou, Li Shen, Jun Yu, Bo Han, Mingming Gong, Nannan Wang, Tongliang Liu

Researchers from The University of Sydney, Xidian University, JD Explore Academy, University of Science and Technology of China, Hong Kong Baptist University, and The University of Melbourne developed Strength-Adaptive Adversarial Training (SAAT), which replaces the fixed perturbation budget in adversarial training with a minimum adversarial loss constraint. This approach dynamically adjusts attack strength to mitigate robust overfitting and maintain a steady robustness disparity across diverse model capacities, achieving up to 50.67% AutoAttack accuracy when combined with AWP on CIFAR10.

Key insights

• Constraining the minimum adversarial loss instead of the maximum perturbation budget allows the attack strength to adapt dynamically to the model's evolving robustness and inherent capacity.
• The parameter ρ provides a direct mechanism to control the degree of robustness disparity and manage the trade-off between natural accuracy and adversarial robustness.
• By continuously challenging the model with adversaries that meet a minimum loss target, SAAT effectively prevents robust overfitting and improves generalization against strong attacks.

Summary by alphaXiv.