Consistency Adversarial Training for Robust Decision Boundaries

Adversarial attacks pose a significant threat to the security of deep learning systems. While Adversarial Training (AT) has emerged as a prominent defense strategy, traditional AT methods focus primarily on training with loss-maximizing adversarial samples, potentially overlooking decision boundary refinement. To address this limitation, we propose Consistency Adversarial Training (CAT), a novel approach that shifts the focus of AT from the endpoint of the attack trajectory to its intersection with the decision boundary. CAT innovates by aligning adversarial samples at decision boundary intersections with their corresponding "ancestor" samples, thereby establishing a more comprehensive defense mechanism. Through extensive experiments and ablation studies, we demonstrate that CAT promotes the formation of a flatter and more precise loss landscape, thereby improving the robustness of neural networks.