面向调制识别的对抗样本研究综述

Review of Adversarial Samples for Modulation Recognition

调制方式识别是认知无线电、电磁对抗等领域中的关键一环,也是进行接收机高效信号处理的重要前提。深度学习具有自主分析、自动特征提取和非线性拟合等传统手段无法比拟的独特优势,其在调制方式识别中表现出了具大潜力,但深度学习模型容易受到对抗样本的攻击,对调制识别任务造成严重影响。尽管对抗样本攻击在计算机视觉、自然语言处理等领域得到了广泛研究,但其在调制识别领域的研究成果较为零散。本文基于调制识别的独特特性,介绍了基于深度学习的调制识别技术,构建了调制识别的问题模型,阐述了目前常见的神经网络在调制识别中的应用现状并列举和对比了调制识别常用数据集及其仿真结果。通过回顾攻击类型、对抗样本生成和防御策略总结了最新的研究成果,建立了不同攻击和防御类别的分类法,并讨论了对抗样本在无线通信中的未来前景。

Modulation recognition is a key component in the fields of cognitive radio,electronic warfare,and other related areas.It is also an important prerequisite for efficient signal processing in receivers.Due to the unique advantages of deep learning,such as autonomous analysis,automatic feature extraction,and nonlinear fitting,which traditional methods cannot match,it has great potential in modulation recognition.However,deep learning models are vulnerable to adversarial attacks,which seriously affect the task of modulation recognition.Although adversarial sample attacks have been widely studied in the fields of computer vision and natural language processing,research results in the field of modulation recognition are relatively scattered.This article introduced the modulation recognition technology based on deep learning,established the problem model of modulation recognition,and elaborated on the application status of common neural networks in modulation recognition,as well as listed and compared commonly used datasets and simulation results of modulation recognition.By reviewing attack types,adversarial sample generation,and defense strategies,we summarized the latest research results,established a classification system for different types of attacks and defence,and discussed the future prospects of adversarial samples in wireless communication.