引用本文
  • 张凤荔,周志远,王瑞锦,黄鑫,韩英军.基于生成对抗网络的无线电数据增扩与分类[J].信息安全学报,2023,8(5):47-60    [点击复制]
  • ZHANG Fengli,ZHOU Zhiyuan,WANG Ruijin,HUANG Xin,HAN Yingjun.Radio Data Augmentation with GAN for Automatic Modulation Classification[J].Journal of Cyber Security,2023,8(5):47-60   [点击复制]
【打印本页】 【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 5814次   下载 4036 本文二维码信息
码上扫一扫!
基于生成对抗网络的无线电数据增扩与分类
张凤荔1, 周志远1, 王瑞锦1, 黄鑫1, 韩英军2
0
(1.电子科技大学信息与软件工程学院 成都 中国 610054;2.四川中烟工业有限责任公司成都卷烟厂 成都 中国 610066)
摘要:
相较于传统的无线电数据特征提取方法,深度学习具有高效灵活的特点,其可以有效提高调制数据识别的性能。然而在实践中,收集大量可靠的无线电调制样本数据有时代价是昂贵和困难的,这在很大程度上限制了深度学习模型的性能。本文提出了基于生成对抗网络(Generative Adversarial Networks,GAN)的无线电调制数据增扩模型RMAbGAN (Radio Modulation dataAugmentation based on Generative Adversarial Networks),该模型通过挖掘不同信噪比与调制方式下的无线电调制数据特征差异,生成符合调制方式与信噪比特点的无线电调制数据,模型中的生成器部分捕获无线电调制数据分布特征,辨识器部分优化生成器性能,两者相互博弈性能不断提升;在此基础上,对无线电数据采样特点与无线电数据传统增强方法进行深度分析与研究,发现了无线电调制数据蕴含的空域特征与时序特征,设计出了能深刻捕获无线电数据空域特征与时序特征的无线电数据分类模型AMCST (Automatic Modulation Classification based Spatial and Temporal feature)。通过大量的对比实验,表明相较于基于旋转变换的无线电调制数据增扩模型,RMAbGAN模型在无线电调制数据增扩方面更具有鲁棒性和泛化能力,可以实现更高的调制分类准确率。此外,相较基于长短期记忆网络(Long Short-Term Memory,LSTM)的调制分类模型、基于残差网络(Residual Networks,ResNet)的调制分类模型等传统模型,AMCST模型在调制数据分类方面更具有稳定性和可用性,同时也具有更高的分类准确率。
关键词:  无线电调制数据增扩|无线电调制数据分类|生成对抗网络|卷积神经网络|循环神经网络
DOI:10.19363/J.cnki.cn10-1380/tn.2023.09.04
投稿时间:2021-12-22修订日期:2022-05-27
基金项目:国家自然科学基金(No. 61802033, No. 61472064, No. 61602096)、四川省区域创新合作项目(No. 2020YFQ0018)、四川省科技计划重点研发项目(No. 2020YFG0475, No. 2018GZ0087, No. 2019YJ0543)、博士后基金项目(No. 2018M643453)、广东省国家重点实验室项目(No. 2017B030314131)、网络与数据安全四川省重点实验室开放课题(No. NDSMS201606)资助.
Radio Data Augmentation with GAN for Automatic Modulation Classification
ZHANG Fengli1, ZHOU Zhiyuan1, WANG Ruijin1, HUANG Xin1, HAN Yingjun2
(1.School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu 610054, China;2.Sichuan China Tobacco Industry Co., Ltd. Chengdu Cigarette Factory, Chengdu 610066, China)
Abstract:
Compared with traditional radio modulation data feature extraction methods, deep learning is more efficient and flexible. So it can effectively improve radio modulation data recognition performance. However, in practice, it is too costly and difficult to collect a vast number of reliable radio modulation data samples that the performance of deep learning models is limited largely. In this paper, we propose a radio modulation data augmentation model named RMAbGAN (Radio Modulation data Augmentation based on Generative Adversarial Networks) based on generative adversarial networks. It can generate the radio modulation data satisfying the characteristics of SNR(Signal Noise Ratio) and modulation type by mining the characteristic differences of radio modulation data under different SNR and modulation types. In our model, the generator captures the distribution characteristics of radio modulation data samples and the discriminator optimizes the generator’s performance and parameters. The performance of the model will continue to improve by learning the generator and discriminator in the process of playing against each other. By the way of deep analyzing and researching radio data sampling methods and traditional radio data enhancement methods, we discover the spatial and temporal features in radio modulation data. According to these potential features, we present a radio modulation data classification model named AMCST (Automatic Modulation Classification based Spatial and Temporal feature) that is able to capture as many potential spatial and temporal attributes as possible in radio modulation data. Through a large number of comparative experiments, it is found that RMAbGAN model has better robustness and generalization ability compared with the radio modulation data enhancement model based on rotation transformation, and can achieve higher accuracy of radio modulation classification. In addition, compared with the traditional radio modulation classification model based on long short-term memory and residual networks, AMCST is so available and stable that can improve classification accuracy.
Key words:  radio modulation data enhancement|radio modulation data classification|generative adversarial networks|convolutional neural networks|recurrent neural networks