抗标签噪声的鲁棒电信诈骗检测方法

打开文本图片集

doi：10.3969/J.ISSN.1672-7274.2025.06.018

中图分类号：TN911.23;TP274;TP3 文献标志码：B 文章编码：1672-7274（2025）06-0053-04

Robust Telecommunication Fraud Detection Method Resistant to Label Noise ZHANGXin，ZHANGSihai

（China Mobile Communications Group Hubei Co.，Ltd.，Wuhan 43oo23，China）

Abstract： The frequent occurrence of telecommunications fraud has brought huge economic losss to society， and traditionalfraud detection methods often perform poorly when faced with label noise.To this end，this paper proposes a new telecommunications fraud detection method （DNW-GCN） that combines graph convolution network （GCN）and dynamic noise sample weighting technology.This method performs feature aggregation through GCNand dynamicallyreduces the weightofnoise samples tosuppressits interferenceon model training.Experiments showthat DNW-GCN cansignificantly improve the detection accuracyand recallrate of malicious samples under diferent noise levels，and still maintains an accuracy rate of more than 96% and a recall rate of 84.7% in a 35% noise environment， demonstrating good performance.Noise immunity and robustness.

Keywords：telecommunications fraud; graph neural network;label noise

1 研究背景

通信产业的迅速崛起在为人们带来便利的同时，也给电信诈骗犯罪提供了可乘之机。（剩余4474字）