Robustness against adversary models on MNIST by Deep-Q Reinforcement Learning based Parallel-GANs

This paper presents a method of enhancing robustness of classification machine learning model. Robustness of computer programming is an important topic, not only affects the security of user information but also stability and performance of program itself. That is, a technique aims to misclassify the machine learning model is called Adversarial Attack. We present an experiment to increase the robustness of machine learning models. The dataset MNIST, which includes hand-writing digits, is used as the experiment subject. Several techniques are applied such as Generative Adversarial Networks in Parallel form, Reinforcement Learning in Deep-Q formation, Dynamic sampling is used as prediction of unknown attack. Black-box is set to be experimental scenario. The experimental results show that the average robustness of the system under several attack conditions is as high as 90%.