Blockchain Anomaly Transaction Detection Method Based on Graph Continual Learning

The rapid growth of blockchain has increased exposure to threats such as phishing, Ponzi schemes, and money laundering, making anomaly detection essential for maintaining platform security. Traditional rule-based methods and static machine learning models often fall short against evolving attack strategies, while existing graph neural network (GNN) approaches, though effective in modeling transaction graphs, require costly retraining and suffer from catastrophic forgetting when new anomalies emerge. These challenges highlight the need for adaptive detection techniques that can evolve alongside adversarial behaviors while preserving knowledge of previously observed patterns. To address this problem, this paper introduces a graph-based continual learning approach for blockchain anomaly detection. The proposed framework incorporates a topology-aware weight preservation module that captures local structural dependencies and stabilizes critical parameters during training. By explicitly modeling topological importance, the method balances knowledge retention and new task adaptation, thereby mitigating catastrophic forgetting. In addition, the framework is designed for modular compatibility with different GNN backbones, requiring only minimal adaptation for specific architectures. This flexibility ensures broad applicability in real-world blockchain systems. Extensive experiments demonstrate that our approach not only achieves high detection accuracy but also maintains stability and robustness across sequential tasks, offering a scalable and effective solution for securing blockchain ecosystems against evolving transaction anomalies.