SORA: Free Second-Order Attacks in Fast Adversarial Training
Researchers have introduced SORA, a novel method for adversarial training (AT) designed to combat catastrophic overfitting in fast AT variants. SORA addresses this by formalizing Epsilon Overfitting (EO) and proposing Perturbation Alignment (PertAlign) to predict overfitting onset. The method dynamically adjusts perturbations based on loss surface geometry, consistently preventing overfitting and achieving state-of-the-art robustness and clean accuracy with improved efficiency. AI
IMPACT Introduces a new technique to improve the robustness and efficiency of AI models against adversarial attacks.