Toward Security Enhancement for Deep Learning-Based Channel Estimation Model Under Adversarial Attacks in Next-G Wireless Networks

The growth of wireless networks where the need to achieve high bit rates, reduce latency, and improve the flexibility of wireless services is the goal for future communication systems, in which channel estimation becomes important to utilize in current and next generations. Due to the need to embrace the Artificial Intelligence (AI) revolution, the need to jump from traditional channel estimation into deep learning (DL)-based channel estimation become essential, especially in 6G wireless networks and beyond, where it shows the leverage performance over the traditional system. However, the main obstacle to the safe process is security. The DL-based models can be facing a serious issue in terms of confidentiality especially when it comes to reliability. This paper investigates the false prediction issue that the DL-based channel estimation model can produce due to the adversarial perturbations on pilot input signals (channel parameters) that threaten the reliability of the model. We aim to develop a Convolutional Neural Network (CNN) model architecture in terms of fetching and utilizing mitigation distillation technique, that could mitigate the consequences of the implemented attacks to maintain the reliability of DL-based channel estimation models in next-generation networks. Simulation results demonstrated that the developed CNN model can be trustworthy up to promising results in predictions comparable to the pre-existed CNN model when perturbed pilot input signals enter the model with the same amount of budget attacks implemented on them.