| 引用本文： |
-
韦锋,胡红爽,孟祥,黄伟庆,魏冬,张巧遇.基于双分支的电磁信号多粒度异常检测技术[J].信息安全学报,已采用 [点击复制]
- wei feng,Hu Hongshuang,Meng Xiang,Huang Weiqing,Wei Dong,Zhang Qiaoyu.Dual-Branch-Based Multi-Granularity Anomaly Detection Technique for Electromagnetic Signals[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 随着无线通信与物联网技术的迅猛发展，电磁设备在各行各业得到了广泛应用。然而，由此带来的安全风险也日益凸显，实时监测电磁信号并及时发现异常成为亟待解决的重要问题。本文提出了一种基于双分支的多粒度异常检测框架，通过在两个分支分别同时分析电磁信号的长时间尺度行为模式与局部频谱细节，实现对不同类型异常的精准识别。在粗粒度检测分支，研究利用Transformer模型预测长时间范围内的信号演化趋势，当预测结果与实际观测值之间出现显著偏差时，判定存在行为层面的异常；在细粒度检测分支，通过时间卷积神经网络的去噪自编码器（TCN-DAE）对信号辐射过程进行重构，当重构误差超出预设阈值时，快速识别短时局部的频谱异常。最终本研究通过加权融合粗细粒度的异常评分，获得全面、精准的综合异常判断结果。基于对讲机信号，电梯电磁泄漏信号和WIFI信号的实测与仿真异常实验验证表明，该框架在干扰、发射过程及频次异常等多种情境下均能保持高检测率与低误报率，尤其在发射过程和频次异常的检测中，本文模型的F1值相比于其他深度学习模型提升了5.12%和7.5%。 |
| 关键词: 电磁信号 异常检测技术 深度学习 多粒度特征 |
| DOI： |
| 投稿时间：2025-05-06修订日期：2025-07-24 |
| 基金项目:国家重点研发计划 |
|
| Dual-Branch-Based Multi-Granularity Anomaly Detection Technique for Electromagnetic Signals |
|
wei feng, Hu Hongshuang, Meng Xiang, Huang Weiqing, Wei Dong, Zhang Qiaoyu
|
| (Institute of Information Engineering，Chinese Academy of Sciences) |
| Abstract: |
| With the rapid advancement of wireless communication and Internet of Things (IoT) technologies, electromagnetic (EM) devices have found widespread applications across numerous industries. However, the resulting increase in EM activity has led to growing security concerns, making real-time monitoring and anomaly detection of EM signals a pressing challenge. This paper proposes a dual-branch, multi-granularity anomaly detection framework designed to accurately identify various types of anomalies by simultaneously analyzing long-term signal behaviors and local spectral characteristics. The coarse-grained branch employs a Transformer-based model to predict the long-term evolution of signal patterns; significant deviations between predictions and actual observations are flagged as behavioral anomalies. Meanwhile, the fine-grained branch utilizes a Temporal Convolutional Network-based Denoising Autoencoder (TCN-DAE) to reconstruct local EM radiation patterns, enabling rapid detection of short-duration spectral anomalies when reconstruction errors exceed a defined threshold. The anomaly scores from both branches are integrated through a weighted fusion mechanism to produce a comprehensive and precise anomaly assessment. Extensive experiments on real and simulated datasets—including walkie-talkie signals, elevator EM leakage, and Wi-Fi signals—demonstrate the framework’s high detection accuracy and low false alarm rates across various scenarios such as interference, abnormal transmission processes, and frequency irregularities. Notably, the proposed model achieves improvements in F1-score of 5.12% and 7.5% for transmission process and frequency anomalies, respectively, compared to existing deep learning approaches. |
| Key words: Electromagnetic signals Anomaly detection technology Deep learning Multi-granularity features |
