A Study on Adversarial Machine Learning in Wireless Communication Systems

Machine learning (ML) has since been a powerful tool in advancing the modeling of many wireless communication systems, e.g., cellular networks, WiFi, vehicular networks, and space-air-ground networks. The benefits of ML have allowed wireless networks to become self-organized, adaptive, and more resilient to random behaviors. However, the availability of synthetic and/or real wireless network data and trained ML model parameters have continued to motivate attackers to exploit their vulnerabilities. Many recent works have shown that ML models lack the ability to be robust against adversarial attacks when an attacker crafts an adversarial example to fool the ML model. One of the key factors is the susceptibility of the air interface of wireless systems to adversarial attacks. The problem, thus, becomes important to be timely studied as adversarial threats may subvert ML models, making their roles in wireless communication insignificant. To clearly understand the ongoing research in this area, this paper explores existing articles that describe vulnerabilities of ML-based models in wireless communication systems, and the defense mechanisms against adversarial attacks. The experiments and results in this paper show an attacker’s exploitation of spectrum sensing data from 100 kHz to 6 GHz.