A Federated Anti-money Laundering Detection Model with Bidirectional Graph Attention Network

Existing anti-money laundering (AML) detection models neglect the importance of capturing topological features and directional information embedded in the transaction graph. Meanwhile, the sharing of financial data is hampered by the fact that it involves the privacy of users and the commercial secrets of organizations, leading to the prevalence of isolated “data silos” in the financial sector. To solve these issues, this paper proposes a federated anti-money laundering detection model with bidirectional graph attention network (FALD-BGAT). It starts by distributing the transaction data to the clients in the form of non-independent homogeneous distribution, and then the client trains the model locally through the two-way graph attention network, after that, the trained model parameters are uploaded to the central server, which completes the aggregation and updating of the model parameters and returns the updated parameters to the clients, and finally, each client starts the next round of training. Compared to existing models, experimental show that the local model of our scheme improves the precision, recall and F1 value by 1%,7%,6% on elliptical dataset and by 4%,17% and 14% in federated mode, respectively.