Studying the Robustness of Anti-Adversarial Federated Learning Models Detecting Cyberattacks in IoT Spectrum Sensors

Device fingerprinting combined with Machine and Deep Learning (ML/DL) report promising performance when detecting spectrum sensing data falsification (SSDF) attacks. However, the amount of needed to train models scenario privacy concerns limit applicability centralized ML/DL. Federated learning (FL) addresses these drawbacks but is vulnerable adversarial participants The literature has proposed countermeasures, more effort required evaluate FL SSDF attacks their robustness against adversaries. Thus, first contribution this work create an FL-oriented dataset modeling behavior resource-constrained sensors affected by second a pool experiments analyzing according i) three families sensors, xmlns:xlink="http://www.w3.org/1999/xlink">ii) eight attacks, xmlns:xlink="http://www.w3.org/1999/xlink">iii) four scenarios dealing anomaly detection binary classification, xmlns:xlink="http://www.w3.org/1999/xlink">iv) up 33% implementing model poisoning xmlns:xlink="http://www.w3.org/1999/xlink">v) aggregation functions acting as anti-adversarial mechanisms. In conclusion, achieves Without mechanisms, are particularly $>$ 16% Coordinate-wise-median best mitigation for detection, classifiers still