融合图增强与多教师动态蒸馏的区块链异常交易检测方法

PDF下载HTML

Title:

A blockchain abnormal transaction detection method integrating graph enhancement and multi-teacher dynamic distillation

作者:

张铭泉^1,2, 贾文莉¹, 孟丽敏³, 赵峻贤¹

1. 华北电力大学 控制与计算机工程学院, 河北 保定 071003;
2. 华北电力大学 复杂能源系统智能计算教育部工程研究中心, 河北 保定 071003;
3. 华北电力大学 网络与信息化办公室, 河北 保定 071003

Author(s):

ZHANG Mingquan^1,2, JIA Wenli¹, MENG Limin³, ZHAO Junxian¹

1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China;
2. Engineering Research Center of Intelligent Computing for Complex Energy Systems, Ministry of Education, North China Electric Power University, Baoding 071003, China;
3. Network and Information Office, North China Electric Power University, Baoding 071003, China

关键词:

区块链; 异常检测; 图神经网络; 知识蒸馏; SMOTE; 多教师模型; 动态蒸馏; 特征增强

Keywords:

blockchain; anomaly detection; graph neural network; dnowledge distillation; SMOTE; multi-teacher model; dynamic distillation; feature enhancement

分类号:

TP309；TP18

DOI:

10.11992/tis.202510002

摘要:

针对现有区块链异常交易检测方法在复杂交易结构建模与数据不平衡处理方面的不足，本文提出一种融合图神经网络多教师模型、motif结构特征、SMOTE(synthetic minority oversampling technique)采样与双学生模型知识蒸馏的检测方法——MGDD (multi-teacher GNN dynamic distillation)。该方法首先利用motif统计量增强节点结构表达，随后结合图卷积网络、图注意力网络、图同构网络、图采样与聚合网络构建多教师模型，通过动态蒸馏机制引导多层感知机与随机森林学生模型学习图中的高阶语义特征，并引入SMOTE进行过采样以缓解样本不均衡问题。在Elliptic数据集上的实验结果表明，MGDD在准确率、召回率与F1值等指标上全面优于现有典型方法，其中RF学生模型在蒸馏引导下准确率达到99.23%。消融与鲁棒性实验进一步验证了motif特征与知识蒸馏机制在提升检测性能与模型稳定性方面的重要作用。综上所述，MGDD为区块链异常交易检测任务提供了一种高效、可推广的建模范式，具备良好的应用前景。

Abstract:

To address the limitations of existing blockchain abnormal transaction detection methods in complex transaction structure modeling and data imbalance handling, this paper proposes a detection method called MGDD (multi-teacher GNN dynamic distillation) that integrates graph neural network teacher models, motif structural features, SMOTE sampling, and multi-student model knowledge distillation. This method first enhances node structural representation using motif statistics, then constructs multiple teacher models by combining GCN, GAT, GIN, and GraphSAGE graph neural networks. Through a dynamic distillation mechanism, it guides multi-layer perceptron (MLP) and random forest (RF) student models to learn high-order semantic features from graphs, and introduces SMOTE for oversampling to alleviate the sample imbalance problem. Experimental results on the Elliptic dataset demonstrate that MGDD outperforms existing typical methods in terms of accuracy, recall rate, and F1-score, with the RF student model achieving an accuracy rate of 99.23% under the guidance of distillation. Ablation and robustness experiments further verify the important role of motif features and knowledge distillation mechanisms in improving detection performance and model stability. In summary, MGDD provides an efficient and scalable modeling paradigm for blockchain abnormal transaction detection tasks, showing promising application prospects.