| 摘要: |
| 针对无人机通信链路中非授信干扰引发的信号异常威胁而现有检测方法存在场景适应性差、标注依赖度高及复杂信号建模不足等问题，提出了基于条件去噪扩散异常检测（Denoising Diffusion Anomaly Detection，DDAD）模型的无人机通信链路异常信号检测方法。通过构建多阶段信号频谱图表征体系，结合扩散模型渐进式去噪机制与概率密度建模优势，通过像素级对比与特征微调策略实现异常定位，完成通信链路异常信号检测任务。实验中通过模拟重放攻击、信号篡改等典型攻击场景，对比端到端（Sequence to Sequence，Seq2Seq）、变分自编码器（Variational Auto-encoder，VAE）及LSTM-GAN等基准方法。结果表明，条件去噪扩散异常检测模型在精确率（98.5%）、召回率（97.2%）和曲线下面积（Area under the Curve，AUC）（0.979）等核心指标上优于对比方法，相较于次优的LSTM-GAN方法，在保持97.8%的F1分数前提下误报率降低2.3%。 |
| 关键词: 无人机 通信链路 异常检测 无监督学习 |
| DOI：10.20079/j.issn.1001-893x.250509002 |
|
| 基金项目:国家自然科学基金资助项目(52402523，U2133203)；天津市高等学校研究生教育改革研究计划项目(TJYG135)；天津市航空装备安全性与适航技术创新中心开放基金(JCZX-2024-KF-01) |
|
| A Method for Detecting Abnormal Signals in UAV Communications Based on Conditional Denoising Diffusion Models |
| WANG Kenian,TIAN Xiaoyi,ZHAO Changxiao |
| (1.Key Laboratory of Civil Aircraft Airworthiness Technology，Tianjin 300300，China;2.School of Safety Science and Engineering，Civil Aviation University of China，Tianjin 300300，China;3.Tianjin Aviation Equipment Safety and Airworthiness Technology Innovation Centre，Tianjin 300300，China) |
| Abstract: |
| In response to the signal abnormal threats caused by non-licensed interference in unmanned aerial vehicle（UAV） communication links，and the problems of poor scene adaptability，high annotation dependence，and insufficient complex signal modeling in existing detection methods，an anomaly signal detection method for UAV communication links based on the conditional denoising diffusion anomaly detection (DDAD) model is proposed.By constructing a multi-stage signal spectrum characterization system，combining the progressive denoising mechanism of the diffusion model and the advantages of probability density modeling，and through pixel-level comparison and feature fine-tuning strategies，the anomaly location is achieved，and the task of detecting abnormal signals in the communication link is completed.Through experiments simulating typical attack scenarios such as replay attacks and signal tampering，the proposed method is compared with benchmark methods such as Sequence to Sequence（Seq2Seq），Variational Auto-encoder（VAE），and LSTM-GAN.The results show that the proposed model outperforms the comparison methods in core indicators such as precision (98.5%)，recall rate (97.2%)，and area under the curve (AUC) (0979).Compared with that of the suboptimal LSTM-GAN method，while maintaining a F1-score of 97.8%，the false alarm rate is reduced by 2.3%. |
| Key words: UAV communication link anomaly detection unsupervised learning |
《电讯技术》编辑部 