Robust Proxy: Improving Adversarial Robustness by Robust Proxy Learning

Recently, it has been widely known that deep neural networks are highly vulnerable and easily broken by adversarial attacks. To mitigate the vulnerability, many defense algorithms have proposed. to improve robustness, works try enhance feature representation imposing more direct supervision on discriminative feature. However, existing approaches lack an understanding of learning adversarially robust representation. In this paper, we propose a novel training framework called Robust Proxy Learning. proposed method, model explicitly learns representations with proxies. end, firstly, demonstrate can generate class-representative features adding class-wise perturbations. Then, use class representative as With features, through proxy framework. Through extensive experiments, verify manually our could increase robustness DNNs.