Performance Evaluation of Physical Attacks against E2E Autoencoder over Rayleigh Fading Channel

The use of Deep Learning (DL) in wireless communication systems is becoming very popular. As an example to the use of DL, the end-to-end (E2E) communication system can be implemented as an autoencoder. However, security and robustness are the main challenges due to the weaknesses of the autoencoders against physical adversarial attacks. Some works have been devoted to addressing these issues using only Additive White Gaussian Noise (AWGN) channel model. In this paper, we investigate the vulnerabilities of autoencoder E2E using a more realistic Rayleigh channel model with fast-fading and slow-fading characteristics studied separately. We apply white-box and black-box adversarial attacks to show in which extent this system is weak against adversaries. We use AWGN channel model as a benchmark to analyze our results. The results show that the adversary has more destructive impacts on the system that involves Rayleigh channel than AWGN channels as it causes a larger block error rate. Also, the black-box attack affects the system of Rayleigh model similar to jamming attacks.