SRC-FD: Secure and Reliable Cryptocurrency Transaction Fraud Detection Under Hybrid Blockchain

With the help of blockchain and federated learning (FL), we can form an anti-fraud alliance involving amounts of cryptocurrency users to detect fraudulent transactions in financial networks. However, there are two challenges when building the cryptocurrency transaction fraud detection model with privacy protection. First, some existing blockchain-enabled FL frameworks would incur high storage space consumption and provide vulnerabilities for attackers to launch information inference attacks. Second, existing adaptive differential privacy (DP) mechanisms applied in FL sacrifice the detection performance when protecting privacy due to they usually ignore the heterogeneous risks of different attacks and the effect of the similarity of gradients on the privacy budget. In response to the above challenges, we design a secure and reliable FL-based cryptocurrency transaction fraud detection method under hybrid blockchain in financial networks, named SRC-FD. For the first challenge, we design a light hybrid blockchain combined with the public blockchain and consortium blockchain to reduce storage consumption and enhance security and reliability. We also propose a hybrid blockchain-based Proof of Scores consensus mechanism according to credibility scores of users to facilitate consensus, named HPoS. For the second challenge, we design adaptive privacy budget and sensitivity calculation methods by considering the similarity of gradients and the heterogeneous risks of attacks. Finally, we evaluate the proposed SRC-FD method in experiments based on two public fraud detection datasets: Bitcoin Heist dataset (Bitcoin) and Elliptic dataset. Compared with the consortium blockchain, SRC-FD method can improve the throughput and reduce memory usage. Compared with three DP methods, SRC-FD can well defend against privacy attacks and achieve the optimal fraud detection performance.