StrGAN for Generating Enhanced Samples

Junfeng Wu, Jinwei Wang, Junjie Zhao, Xiangyang Luo

Abstract: Most studies use the negative effect of perturbation to mislead the neural network, such as adversarial examples, while ignoring the positive effect of improving neural networks’ performance. In this work, we use enhanced samples with positive perturbation to improve target classifiers’ performance and propose an algorithm of strong generative adversarial networks (StrGAN) to generate enhanced samples. StrGAN directly generates enhanced samples of unlabeled data. Since StrGAN and the target classifier are independent of each other, it can effectively reduce the classifier’s computing resources and training time while improving the performance. The experiment shows that the enhanced samples generated by StrGAN have higher accuracy than original samples, and its accuracy can increase by up to 28.6%.