FMTD: Federated Learning-Based Multi-Angle Feature Fusion Framework for Abnormal Transaction Detection in Digital Currency

With the rapid development of digital currencies, their anonymity has provided shelter for criminals, leading to an increasing number of abnormal transactions that threaten financial order. In response, researchers have proposed centralized methods and federated learning to detect abnormal transactions. However, existing methods suffer from insufficient feature fusion, leading to information loss. In addition, the averaging of client parameters in federated learning negatively affects transaction detection effectiveness. To tackle these issues, we propose a Federated learning-based Multi-angle feature fusion framework for abnormal Transactions Detection, named FMTD. First, an Adaptive Multi-angle Contrastive Learning model (AMCL) is presented as the local model for clients, which improves feature fusion by maximizing positive sample consistency, minimizing negative sample consistency, and assigning higher weights to more effective features. Second, a Federated adaptive algorithm with Adjustable Client Weights (Fed-ACW) is introduced, which reduces the negative impact of poor local models on the global model. While preserving transaction feature privacy, FMTD enhances the transaction detection rate. Experimental results demonstrate its superiority over existing methods.