Machine Learning–Driven Detection of Suspicious On-Chain Activity for Strengthening Security in Blockchain-Based Financial Transactions
Year: 2025
Type: article
Abstract: Blockchain-based financial transactions face increasing risks from illicit on-chain activity, requiring adaptive analytical systems capable of identifying complex anomalies beyond conventional rule-based mechanisms. This study aims to evaluate a machine-learning-driven framework for detecting suspicious on-chain activity using graph-temporal features to strengthen blockchain financial security through accuracy comparisons, statistical validation, and model-specific performance estimation. A quantitative analytical evaluation was conducted at the Department of Marketing & Business Analytics, East Texas A&M University (January–December 2024). A purposive sample of 42 blockchain transaction datasets was processed using supervised classifiers, unsupervised anomaly detectors, and graph neural networks. Performance was assessed using accuracy, precision, recall, F1-score, ROC-AUC, confusion matrices, class-imbalance handling, cross-validation, and statistical significance testing using independent t-tests. The graph-neural-network model achieved the highest detection accuracy (91.7% ± 2.4 SD), outperforming supervised classifiers (85.3% ± 3.1 SD) and anomaly detectors (79.6% ± 4.2 SD). Mean precision for illicit-transaction identification was 89.2%, recall 90.5%, and F1-score 89.8%. A statistically significant difference was observed across model groups (ANOVA, p = 0.003). Additional variables showed strong predictive contribution: multi-hop transaction depth (β = 0.41, p = 0.012), temporal burst intensity (β = 0.37, p = 0.021), address-centrality score (β = 0.33, p = 0.028), and contract-interaction entropy (β = 0.29, p = 0.036). ROC-AUC for the final ensemble averaged 0.94. Machine-learning-enhanced graph-temporal modelling demonstrates strong potential for detecting suspicious blockchain activity. Integrating multilayer features significantly improves security analytics and provides a scalable foundation for advanced financial-risk monitoring.
Author: Asadul Arifin Shawn
Institution: East Texas A&M University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416765607