Detecting Adversarial Examples via Prediction Difference for Deep Neural Networks

Deep neural networks (DNNs) perform effectively in many computer vision tasks. However, DNNs are found to be vulnerable to adversarial examples which are generated by adding imperceptible perturbations to original images. To address this problem, we propose a novel defense method, transferability prediction difference (TPD), to drastically improve the adversarial robustness of DNNs with small sacrificing verified accuracy. We find out that the adversarial examples have lager prediction difference for various DNN models due to their various complicated decision boundaries, which can be used to identify the adversarial examples by converging decision boundaries to a prediction difference threshold. We adopt the K-means clustering algorithm on benign data to determine transferability prediction difference threshold, by which we can detect adversarial examples accurately and efficiently. Furthermore, TPD method neither modifies the target model nor needs to take knowledge of adversarial attacks. We perform four state-of-the-art adversarial attacks (FGSM, BIM, JSMA and C&W) to evaluate TPD models trained on MNIST and CIFAR-10 and the average detection accuracy is 96.74% and 86.61%. The results show that TPD model has high detection ratio on the demonstrably advanced white-box adversarial examples while keeping low false positive rate on benign examples.