One-Pixel Adversarial Example that is Safe for Friendly Deep Neural Networks

Abstract

Deep neural networks (DNNs) offer superior performance in machine learning tasks such as image recognition, speech recognition, pattern analysis, and intrusion detection. In this paper, we propose a one-pixel adversarial example that is safe for friendly deep neural networks. By modifying only one pixel, our proposed method generates a one-pixel-safe adversarial example that can be misclassified by an enemy classifier and correctly classified by a friendly classifier. To verify the performance of the proposed method, we used the CIFAR-10 dataset, ResNet model classifiers, and the Tensorflow library in our experiments. Results show that the proposed method modified only one pixel to achieve success rates of 13.5% and 26.0% in targeted and untargeted attacks, respectively. The success rate is slightly lower than that of the conventional one-pixel method, which has success rates of 15% and 33.5% in targeted and untargeted attacks, respectively; however, this method protects 100% of the friendly classifiers. In addition, if the proposed method modifies five pixels, this method can achieve success rates of 20.5% and 52.0% in targeted and untargeted attacks, respectively.