An MCTS-based Adversarial Training Method for Image Recognition

Yi-Ling Liu, Alessio Lomuscio

Abstract: We present an adversarial training algorithm based on Monte Carlo Tree Search. We illustrate the robustness of the algorithm by studying its resistance to adversarial examples in the context of the MNIST and CIFAR10 datasets. For MNIST, after 2000 epochs the experimental results showed an average improvement of efficiency of 21.1% when compared to PGD. For CIFAR10, after 7000 epochs we obtained an average improvement of efficiency of 9.8% compared to PGD. We further compare the robustness of the algorithm against previous work against various attack methods. The results suggest that the adversarial training method here introduced is not only robust with respect to adversarial examples but also efficient during training.