Stochastic Quantized Activation: To prevent Overfitting in Fast Adversarial Training

Existing neural networks are vulnerable to "adversarial examples"---created by adding maliciously designed small perturbations in inputs to induce a misclassification by the networks. The most investigated defense strategy is adversarial training which augments training data with adversarial examples. However, applying single-step adversaries in adversarial training does not support the robustness of the networks, instead, they will even make the networks to be overfitted. In contrast to the single-step, multi-step training results in the state-of-the-art performance on MNIST and CIFAR10, yet it needs a massive amount of time. Therefore, we propose a method, Stochastic Quantized Activation (SQA) that solves overfitting problems in single-step adversarial training and fastly achieves the robustness comparable to the multi-step. SQA attenuates the adversarial effects by providing random selectivity to activation functions and allows the network to learn robustness with only single-step training. Throughout the experiment, our method demonstrates the state-of-the-art robustness against one of the strongest white-box attacks as PGD training, but with much less computational cost. Finally, we visualize the learning process of the network with SQA to handle strong adversaries, which is different from existing methods.