Persistent Clean-Label Backdoor on Graph-Based Semi-supervised Cybercrime Detection

Cybercrime, which involves the use of tactics such as hacking, malware attacks, identity theft, ransomware, and online scams, has emerged as a major concern for public security management recently. To combat massive cybercrime and conduct a clean Internet environment, graph-based semi-supervised cybercrime detection (GSCD) has gained increasing popularity recently for it can model complex relationships between network objects and provide node-level behavior predictions. However, in this paper, we present a novel threat on GSCD, named clean-label backdoor attack on GSCD (CBAG), which may be utilized by attackers to escape cybercrime detection successfully. The CBAG patches node features of unmarked training data with adversarially-perturbed triggers to enforce the well-trained GSCD model to misclassify trigger-embedded crime data as the premeditated result. Extensive experiments on multiple detection models and open-source datasets reveal that the CBAG exhibits effective escape performance and evasiveness.