Amata: An Annealing Mechanism for Adversarial Training Acceleration

Despite the empirical success in various domains, it has been revealed that deep neural networks are vulnerable to maliciously perturbed input data much degrade their performance. This is known as adversarial attacks. To counter attacks, training formulated a form of robust optimization demonstrated be effective. However, conducting brings computational overhead compared with standard training. In order reduce cost, we propose an annealing mechanism, Amata, associated The proposed Amata provably convergent, well-motivated from lens optimal control theory and can combined existing acceleration methods further enhance It on datasets, achieve similar or better robustness around 1/3 1/2 time traditional methods. addition, incorporated into other algorithms (e.g. YOPO, Free, Fast, ATTA), which leads reduction large-scale problems.