MF-BFDSA: A Multi-Feature Fusion and Dynamic Weight Algorithm for Real-Time Blockchain Financial Transaction Risk Detection

Abstract

To address the challenges of insufficient accuracy in identifying complex risks in blockchain financial transactions and high detection latency in high-concurrency scenarios, this paper constructs a multi-feature fusion dynamic security detection algorithm (MF-BFDSA) and designs a security enhancement mechanism. This algorithm innovatively integrates three types of features: transaction data (amount, frequency, etc.), user behavior, and blockchain network (node response latency, etc.). It uses Q-learning to dynamically adjust weights and combines an improved LSTM (using a forget gate to introduce feature correlation coefficients) with an optimized XGBoost algorithm to construct a risk identification model. This mechanism utilizes a five-layer architecture encompassing data collection and algorithm detection, forming a dual barrier of "algorithm pre-detection + blockchain strong verification." Experiments were conducted on the Hyperledger Fabric 2.4 platform, using Kaggle credit card fraud (284,000 records), IEEE-CIS financial fraud (550,000 records), and 1.05 million simulated transaction data. Comparing traditional algorithms with state-of-the-art methods, the results show that MF-BFDSA achieves an accuracy of 99.82% (surpassing state-of-the-art by 1.2-3.5 percentage points), an F1 score of 98.94% (surpassing state-of-the-art by 2.8-4.2 percentage points), and an anomaly detection latency of only 0.08 seconds. It also maintains a throughput of 1200 TPS and a latency of 0.15 seconds on mid-range hardware. A blockchain system integrating this algorithm achieves a throughput of 1800 TPS (50% higher than the original system) at 2000 TPS concurrency, a CPU utilization of 72% (reduced by 13 percentage points), and a double-spend success rate of less than 0.001%, validating the effectiveness of the algorithm and mechanism.