基于输入通道拆分的无线通信网络对抗攻击多任务防御

Wireless communication network counter attack multi-task defensebased on input channel splitting

在无线通信网络中,由于网络的开放性和共享性,攻击源自多个不同的源头,展现出多种多样的特征。传统的防御方法难以同时应对多种攻击模式,且在处理多模态数据时存在效率低下和准确性不足的问题。为此,研究基于输入通道拆分的无线通信网络对抗攻击多任务防御方法。利用Morlet小波变换将无线通信网络信号转换为时频图像,以输入通道拆分的方式拆分时频图像,得到RGB三个通道的时频图像。在改进注意力机制生成对抗网络内,结合多任务学习建立多防御模型。该模型内生成器通过空间注意力模块与时间注意力长短期记忆网络模块,提取RGB三个通道时频图像的时空特征,并生成对抗样本,通过判别器识别图像类型。检测到攻击时,用对抗样本替换攻击数据,实现无线通信网络的多任务对抗防御。实验证明,该方法可有效将无线通信网络信号转换成时频图像,且有效生成对抗样本,完成无线通道网络对抗攻击多任务防御。

In wireless communication networks,attacks come from various sources and exhibit a variety of characteristics because of the openness and sharing of the network.It is difficult for the traditional defense methods to deal with multiple attack modes simultaneously.In addition,the defense efficiency is low and the accuracy is unsatisfied when processing multimodal data.Therefore,a wireless communication network counter attack multi-task defense based on input channel splitting is proposed.The Morlet wavelet transform is used to convert wireless communication network signals into time-frequency images,the time-frequency images are split into three RGB channels by input channel splitting.By the improved self-attention generative adversarial network(SAGAN),a multi-defense model is established in combination with multi-task learning.The generator within this model extracts the spatiotemporal features of RGB channel time-frequency images by the spatial attention module and the temporal attention long short-term memory(LSTM)network module,and generates adversarial samples.The image types are identified by discriminators.When an attack is detected,the attack data is replaced with adversarial samples to achieve multi-task adversarial defense in wireless communication networks.Experimental results have shown that the proposed method can effectively convert wireless communication network signals into time-frequency images and generate adversarial samples effectively,so as to complete wireless communication network counter attack multi-task defense.