Multiuser Adversarial Attack on Deep Learning for OFDM Detection

Abstract

Adversarial attack has been widely used to degrade the performance of deep learning (DL), especially in the field of communications. In this letter, we evaluate different whitebox and black-box adversarial attack algorithms for a DL-based multiuser orthogonal frequency division multiplexing (OFDM) detector subject to multiuser adversarial attack. The bit error rates under different adversarial attacks are compared. The results show that, the perturbation efficiency of adversarial attack is higher than conventional multiuser interference. Virtual adversarial methods (VAM) and zeroth-order-optimization (ZOO) attacks perform the best among white-box and black-box methods, respectively. They are also effective when the attack changes the starting time. Additionally, adding the number of attackers is found useful to improve the VAM attack but not for ZOO. This work shows that adversarial attack is powerful to generate adversarial against multiuser OFDM communications.