L Nguyen, S Wang, and A Sinha
Ninth Conference on Decision and Game Theory for Security, October 2018.
Deep Neural Networks (DNNs) have been shown to be vulnerable against adversarial examples, which are data points cleverly constructed to fool the classifier. In this paper, we introduce a new perspective on the problem. We do so by first defining robustness of a classifier to adversarial exploitation. Further, we categorize attacks in literature into high and low perturbation attacks. Next, we show that the defense problem can be posed as a learning problem itself and find that this approach effective against high perturbation attacks. For low perturbation attacks, we present a classifier boundary masking method that uses noise to randomly shift the classifier boundary at runtime. We also show that both our learning and masking based defense can work simultaneously to protect against multiple attacks. We demonstrate the efficacy of our techniques by experimenting with the MNIST and CIFAR-10 datasets.