面向智能调制识别的电磁信号灵巧诱骗方法

Radio Signal Smart Deception Method for Intelligent Modulation Classification

基于深度学习的调制识别技术以其在特征提取和识别性能方面的优势,在电磁频谱感知中逐渐得到融合应用。然而,由于智能频谱感知模型内在学习机制的脆弱性和局限性,通过在原始信号上添加难以感知的扰动,能够降低电磁信号调制识别系统的性能。本文针对智能电磁信号调制识别系统的脆弱性,开展了电磁信号诱骗方法研究。与传统的在信号上添加全局扰动不同,该方法通过优选原始电磁信号上的显著采样点,并在特定采样点上加入扰动,实现对电磁信号调制识别系统的灵巧诱骗。实验结果表明,该方法通过在原始信号上添加轻微的隐蔽扰动,在不影响接收机性能的前提下,即可使得智能调制识别系统产生机器幻觉,将对抗样本识别为指定的信号类型,定向扰乱电磁信号调制识别系统的识别结果,从而达到“隐真示假”的目的。

Modulation classification techniques based deep learning are gradually being integrated and applied with their advantages in feature extraction and recognition performance. However, due to the vulnerability and limitation of the intelligent spectrum sensing model, the performance can be reduced by adding imperceptible perturbation to the original signal. Aiming at the vulnerability of intelligent electromagnetic signal modulation classification system, this paper carried out the research of electromagnetic signal deception method. Different from adding global disturbance on the signal, this method added disturbance to the specific sampling points, and the electromagnetic signal modulation classification system can be deceived smartly. The results show that this method could produce a relatively slight invisible disturbance, which deceives the system produce machine hallucinations and identify adversarial example as targeted signal and disturb the identification result of the system, so as to achieve the goal of “concealing the true and revealing the false”.