摘要
基于深度学习的频谱感知模型虽检测性能优异,但普遍具有脆弱性,容易受到频谱对抗攻击的干扰。为了防御这种攻击,提出使用降噪自编码器过滤对抗信号,并在此基础上提出了一种结合降噪自编码器和防御蒸馏的联合防御方法。利用对抗信号和干净信号预训练得到降噪自编码器,频谱信号经降噪自编码器过滤后用于训练感知分类器,在测试阶段,联合使用降噪自编码器和分类器。同时,为进一步缓解扰动对模型性能的影响,在分类器训练阶段,利用蒸馏算法平滑训练网络,提高模型泛化能力。实验结果表明,对于可以降低深度学习频谱感知模型检测概率的频谱对抗攻击,所提出的基于降噪自编码器的防御方法仍然能够拥有较高的检测概率和较低的攻击成功率。
Although the spectrum sensing model based on deep learning has excellent detection performance,it is generally vulnerable to the interference from spectrum adversarial attacks.To defend against such attacks,this paper proposes to filter the adversarial signals using denoising autoencoders.Based on this,a joint defense method is proposed combining denoising autoencoders and defensive distillation.The denoising autoencoder is obtained by pre-training the adversarial signal and clean signal,and after being filtered by the denoising autoencoder,the spectral signal is used to train the spectrum sensing classifier.Then in the testing pahse,the denoising autoencoder and classifier are jointly used.Meanwhile,in order to further alleviate the impact of perturbation on the model performance,a distillation algorithm is used to smooth the training network in the classifier training phase to improve the corresponding generalization ability.The experimental results show that the proposed defense method based on a denoising autoencoder can still have a high detection probability and low attack success rate for deep learning spectrum sensing models whose detection probability is reduced by spectrum adversarial attacks.
作者
杨研蝶
李志刚
张思成
包志达
林云
YANG Yandie;LI Zhigang;ZHANG Sicheng;BAO Zhida;LIN Yun(School of Information and Communication Engineering,Harbin Engineering University Harbin 150006,China)
出处
《移动通信》
2023年第2期28-36,共9页
Mobile Communications
基金
中央高校基本科研业务费资助项目(3072022CF0804)。
关键词
频谱感知
对抗攻击
降噪自编码器
防御蒸馏
spectrum sensing
adversarial attacks
denoising autoencoder
defensive distillation
