AP-GCL: Adversarial Perturbation on Graph Contrastive Learning

Abstract

A serious ecological hazard of illegal transactions (money laundering, financial fraud, etc.) on the Bitcoin trading network. Anti-money laundering and fraud detection are essential instruments to address the problem. However, such datasets are generally extremely unbalanced in terms of positive and negative samples, and most of the data are unlabelled, with the illegal class accounting for just a minimal fraction of the total, which prevents supervised learning from learning a well-represented feature. We propose a self-supervised learning framework based on contrastive learning, in which two different augmented transformations are applied to the original graph data, perturbations are randomly attached to the node features of the upgraded views, and the model parameters and perturbations are updated by gradient descent to maximize the consistency of the single node in different views. The experimental result demonstrates that our model achieves excellent performance in all metrics and is comparable to supervised methods, which verifies the efficiency of the perturbation-based contrastive learning model.