Improving robustness of convolutional neural networks using element-wise activation scaling

Recent works reveal that re-calibrating intermediate activation of adversarial examples can improve the robustness CNN models. The state arts exploit this feature at channel level to help models defend attacks, where each is uniformly scaled by a factor. However, we conduct more fine-grained analysis on and observe only change portion elements within an activation. This observation motivates us investigate new method re-calibrate CNNs robustness. Instead scaling activation, individually adjust element thus propose Element-Wise Activation Scaling, dubbed EWAS, CNNs’ EWAS simple yet very effective in enhancing Experimental results ResNet-18 WideResNet with CIFAR10 SVHN show significantly improves accuracy. Especially for ResNet18 CIFAR10, increases accuracy 37.65% 82.35% against C&W attack. code trained are available https://github.com/ieslab-ynu/EWAS.