RNAS: Robust Network Architecture Search beyond DARTS
The vulnerability of Deep Neural Networks (DNNs) (i.e., susceptibility to adversarial attacks) severely limits the application of DNNs. Most of the existing methods improve the robustness of the model from weights optimization, such as adversarial training and regularization. However, the architecture of DNNs is also a key factor to robustness, which is often neglected or underestimated. We propose a Robust Network Architecture Search (RNAS) to address this problem. In our method, we define a network vulnerability metric based on the features’ deviation between clean examples and adversarial examples. Through constraining this vulnerability, we search the robust architecture and solve it by iterative optimization. The extensive experiments conducted on CIFAR-10/100 and SVHN show that our model achieves the best performance under various adversarial attacks compared with extensive baselines and state-of-the-art methods.
PDF Abstract