A privacy-preserving federated deep learning framework for real-time fraud detection in decentralized financial systems

Abstract

The rapid growth of decentralized financial systems has introduced sophisticated fraud that conventional detection methods struggle to identify, due to fragmented data, strict privacy regulations, and the need for real-time response. This study proposes Federated Differentially Private Fraud Detection (FedDP-FD), a privacy-preserving federated deep learning framework that combines differential privacy with a hybrid long short-term memory graph neural network (LSTM-GNN) for real-time fraud detection in decentralized financial networks. To the best of our knowledge, this is among the first works to jointly optimize collaborative privacy-preserving training, temporal-relational fraud patterns, and low-latency inference under the high-throughput and heterogeneous conditions of these networks. Each participant trains a hybrid model locally on private transaction data and shares only differentially private gradients. FedDP-FD employs two bidirectional long short-term memory (LSTM) layers and two GraphSAGE layers whose embeddings are fused via concatenation and dense layers. With adaptive gradient clipping and Gaussian noise, the framework achieves formal sample-level differential privacy of ( ϵ , δ ) ≈ ( 8.48,1 0 − 5 ) . Knowledge distillation, magnitude-based pruning (30%–40%), and 8-bit quantization enable micro-batch inference with early exits. Experiments on Ethereum, Elliptic, and synthetic datasets yield accuracies of 92.9%–94.5%, F1-scores of 0.637–0.739, and precision–recall area under the curve (PR-AUC) values of 0.695–0.804. Given the severe class imbalance (fraud ratio 1.9%–2.2%), the PR-AUC and low false-positive rate demonstrate practical utility. FedDP-FD thus offers a strong trade-off among detection performance, privacy protection, and computational cost, with potential extensions to healthcare and the Internet of Things (IoT).