Research On Attack And Defense Methods For Intelligent Channel Estimation

Channel estimation can obtain the channel state according to the received signal information,and then obtaining the impulse response of the channel,thereby recovering the original signal.How to accurately perform channel estimation with low overhead is one of the key issues to improve the performance of communication systems.In recent years,with the development of deep learning technology,channel estimation technology based on deep networks has been widely studied,which effectively solves the problems of non-intelligent channel estimation algorithms that require a priori channel statistics,large pilot overhead,and poor robustness.However,studies have shown that deep networks are highly susceptible to adversarial attacks because of their differentiability.Attackers only need to design small noise to add to the original data to make the output of the network change significantly.In practical applications,both the security and robustness of intelligent channel estimation have extremely high requirements.However,there are few researches on adversarial attacks in this field at present.In order to explore the vulnerability and security blind spots of intelligent channel estimation algorithms,this thesis will carry out research on adversarial attack methods,and on this basis,carry out research on defense methods to enhance intelligence security and robustness of channel estimation methods,the main work is as follows:1.Adversarial attack methods against a single signal are studied.By combining classical attack methods such as FGSM,BIM,MIM,and CW with signal data features,an adversarial attack signal for intelligent channel estimation is generated,and an effective attack on intelligent channel estimation is realized.2.A general adversarial attack method is studied.Through the iterative calculation of the signal data set,a disturbance signal that can attack most signals is generated,which improves the real-time problem of the single-signal attack method.The general-purpose adversarial attack method has great advantages in concealment and efficiency,but It is weaker than the single-signal attack method in terms of migration ability.3.A defense method based on adversarial training is studied.From the perspective of model defense,the projected gradient descent method is used to generate the attack signal,and the signal is used to train the network.The generated network model can effectively reduce the bit error rate of channel estimation under adversarial attack.4.A defense method based on denoising autoencoders is investigated.From the perspective of signal data defense,the noise reduction autoencoder is trained,which improves the problem of high model training overhead in the adversarial training method,and reduces the bit error rate of channel estimation after the attack signal passes through the autoencoder.