Black-box Adversarial Examples against Intelligent Beamforming in 5G Networks

Due to the shared and open nature of wireless medium, multiple essential components of the next-generation communication systems that rely on artificial intelligence are vulnerable to adversarial example generation attacks. The goal of these attacks is to manipulate the inputs to the machine learning models during the inference stage over the air in such a way that when the perturbed samples are fed to the target model it returns wrong outputs. This may compromise the integrity of the target model and negatively affect the functionality of the corresponding network framework under attack. In this study, we focus on adversarial example generation attacks against machine learning models which can be deployed in future 5G networks for intelligent beamforming. First, we discuss applications of machine learning for beam selection as well as vulnerability of the resulting intelligent frameworks to adversarial examples. After that, we evaluate various input-dependent black-box adversarial example attacks for multiple beamforming use cases based on deterioration of the network throughput. Finally, we implement and test several algorithms for crafting universal input-agnostic adversarial perturbations which can be employed by adversaries in realistic attack scenarios.