DADF: A Dynamic Adaptive Method for Generating Adversarial Examples

Deep neural networks (DNNs) have made remarkable achievements in several areas, such as image classification. However, extensive researches show that DNNs are vulnerable to adversarial examples which are generated by adding some small perturbations on original images. DeepFool is an effective algorithm for generating adversarial examples with higher success rate and smaller perturbations. However, DeepFool computes the distance of all classification boundaries of the dataset, which causes the speed of DeepFool to be very slow. To address this problem, this paper proposes DADF, a dynamic adaptive method for generating adversarial examples. Based on tracing the information of the attacking process, DADF dynamically selects the subset of target labels. The subset of target labels is dynamically updated once successful adversarial examples are generated. Experimental results on MNIST and CIFAR10 datasets show that, comparing with DeepFool, DADF reduce the time greatly while generating 10000 adversarial examples on the MNIST and CIFAR10 dataset.