A General Framework for Defending against Backdoor Attacks via Influence Graph

Xiaofei Sun, Jiwei Li, Tianwei Zhang, Han Qiu, Fei Wu, Chun Fan

Abstract: In this work, we propose a new and general framework to defend against backdoor attacks, inspired by the fact that attack triggers usually follow a \textsc{specific} type of attacking pattern, and therefore, they have greater connections with and larger impacts on each other during training. We introduce the notion of the {\it influence graph}, which consists of nodes and edges respectively representative of individual training points and associated pair-wise influences. The influence between a pair of training points represents the degree to which removing one training point would impact the prediction the model makes on the other training point, approximated by influence function. Malicious training points are extracted by finding the maximum average sub-graph subject to a particular size. Extensive experiments on computer vision and natural language processing tasks demonstrate the effectiveness and generality of the proposed framework.