Exploring Abnormal Behavior in Swarm: Identify User Using Adversarial Examples

With appearing of adversarial examples which are able to fool deep neural networks, some defense methods are developed to recognize adversarial images. This paper focuses on a new kind of detection on adversarial examples, which aims to find the abnormal user who utilized adversarial examples among a number of normal users. Due to the utilization of adversarial images, the abnormal user is significantly deviated from the majority normal users. Based on this, we propose a straightforward method to extract abnormity information, and design a pooled detection scheme to identify the abnormal user by hierarchical clustering. Experimental results show that our scheme is able to identify the utilization of popular adversarial example methods, and achieve low computational complexity.