Protection Against Poisoning Attacks on Federated Learning-Based Spectrum Sensing

Federated-Learning (FL) based Spectrum Sensing (SS) method is considered for the application in future cognitive radio communication systems due to its supreme performance in changing radio environments as compared to classic cooperative or non-cooperative SS. It also avoids transferring large training datasets with high-resolution localization data. The FL algorithm is the subject of poisoning attacks that can be random or coordinated. In this paper, we first evaluate the impact of such attacks on the FL-based SS performance. Next, we propose a zero-trust method based on continuous monitoring and classification of the sensors’ models to detect attacked models. These models are then eliminated from the global model construction in FL. Our method is <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">semi-blind</i>, i.e., it does not require an apriori knowledge of who are the genuine actors participating in FL. Simulation results of the system under various attacks (random or coordinated, moderate or very aggressive, deliberately increasing or decreasing the spectrum occupancy) show that our method decreases the SS probability of false alarms by 89 % and increases the SS probability of detection by 16 % in case of the most severe targeted attacks in the most critical SNR ranges.