Federated Deep Learning for Privacy-Preserving Cryptocurrency Fraud Detection Using Heterogeneous Financial Data
TL;DR
FedHDL (Federated Heterogeneous Deep Learning), a novel privacy-preserving framework for cryptocurrency fraud detection that enables collaborative model training across five heterogeneous institutional nodes without raw data exchange, is introduced.
Abstract
Cryptocurrency fraud encompassing pump-and-dump schemes, Ponzi contracts, phishing, ransomware laundering, and exchange manipulation inflicted estimated losses of USD 9.9 billion globally in 2023. Existing centralised fraud detection systems require pooling sensitive transactional data across financial institutions and blockchain analytics firms, posing severe privacy, regulatory, and competitive risks. This paper introduces FedHDL (Federated Heterogeneous Deep Learning), a novel privacy-preserving framework for cryptocurrency fraud detection that enables collaborative model training across five heterogeneous institutional nodes without raw data exchange. FedHDL integrates a Graph Attention Network (GAT) for transaction graph encoding, a Bidirectional LSTM (BiLSTM) for temporal behavioural modelling, and a Transformer-based attention fusion layer to reconcile divergent feature spaces across nodes with non-IID (non-independent and identically distributed) data distributions. Model aggregation employs a novel Reputation-Weighted Federated Averaging (RW-FedAvg) algorithm resistant to Byzantine gradient attacks. Privacy is enforced through the synergistic application of Rényi Differential Privacy (RDP) with a Gaussian mechanism (ε = 2.0, δ = 10⁻⁵) and additive homomorphic encryption of gradient updates. Evaluated on 276.23 million real-world and synthetic cryptocurrency transactions spanning five institutional nodes, FedHDL achieves an F1-score of 94.4%, AUC-ROC of 0.981, and accuracy of 96.4% surpassing the best competing federated baseline (FedProx+LSTM) by 7.7 percentage points in F1 and approaching centralised oracle performance (F1 = 95.1%) within 0.7 percentage points, while preserving strict data locality. Communication overhead is reduced by 89.7% relative to uncompressed FedAvg through gradient sparsification and top-k compression. These results demonstrate that FedHDL constitutes a practical, privacy-compliant, and high-fidelity solution for cross-institutional cryptocurrency fraud detection.