MRobust: A Method for Robustness against Adversarial Attacks on Deep Neural Networks

We present a novel black-box adversarial training algorithm to defend against state-of-the-art attack methods in machine learning. In order to search for an adversarial attack, the algorithm analyses small regions around the input that are likely to make significant contributions for the generation of adversarial samples. Unlike some of the literature in the area, the proposed method does not require access to the internal layers of the model and is therefore applicable to applications such as security. We report the experimental results obtained on models of different sizes built for the MNIST and CIFAR10 datasets. The results suggest that known attacks on the resulting models are less transferable than those models trained by state-of-the art attack algorithms.