基于Stacking集成学习的区块链异常交易检测技术研究  

Research on Blockchain Anomaly Transaction Detection Technology Based on Stacking Ensemble Learning

作　　者：王志强[1] 王姿旖 倪安发 Wang Zhiqiang;Wang Ziyi;Ni Anfa(Beijing Electronics Science&Technology Institute,Beijing 102627)

机构地区：[1]北京电子科技学院,北京102627

出　　处：《信息安全研究》2023年第2期98-108,共11页Journal of Information Security Research

基　　金：国家信息中心重点研发计划项目(2018YFB0803401);中国博士后科研基金项目(2019M650606);北京电子科技学院一级学科建设项目(3201012)。

摘　　要：为了高效地进行区块链异常交易检测,提出了一种基于Stacking集成学习的区块链异常交易检测方法.首先,采用XGBoost, LightGBM,CatBoost, LCE作为基分类器,采用MLP作为元分类器,设计了MLP_Stacking集成学习算法;其次,利用SUNDO进行数据扩充,解决数据集中严重类不均衡问题;最后,设计多模型联合特征排序算法,生成最优特征子集,将得到的最优特征子集作为MLP_Stacking输入数据集进行分类训练,通过网格搜索优化参数实现模型优化.实验采用Kaggle平台提供的开源数据集,实验结果显示采用SUNDO数据生成方法能有效提高各分类器性能,在此基础上,设计的集成模型训练效果明显优于单个模型.In order to efficiently detect abnormal transactions on the blockchain, this paper proposes a method based on Stacking integration learning. Firstly, XGBoost, LightGBM, CatBoost and LCE are used as the base classifier, and MLP is used as the metaclassifier, and the MLP_Stacking integrated learning algorithm is designed. Secondly, SUNDO is used for data augmentation to solve the problem of serious imbalance in data sets;Finally, a multi-model joint feature sorting algorithm is designed to generate an optimal subset of features, and the resulting optimal subset of features is used as the input data set of the MLP_Stacking for classification training to achieve model optimization. This paper experiments at the open source dataset provided by Kaggle platform, and the experimental results show that the SUNDO data generation method can effectively improve the performance of each classifier, and the training effect of the integrated model designed in this paper is obviously better than that of the individual model.

关 键 词：类不平衡 STACKING 异常检测 区块链 集成学习 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]