DBFFL - Defending Against Dynamic Poisoning Attacks in Federated Learning in 5G and Beyond Systems

Federated Learning (FL) is a distributed Machine Learning (ML) technique that trains a collaborative ML model without sharing data and by sharing only the model updates. FL is critical in 6 G networks as it enables decentralized and privacy-preserving Artificial Intelligence (AI) model training across distributed devices, reducing communication overhead and enhancing data security, which is crucial for supporting AI-driven 6G networks and applications such as autonomous vehicles, healthcare, and immersive Extended Reality. Since the FL clients transmit only the model updates instead of data, FL is vulnerable to poisoning attacks. The result of a poisoning attack is an incorrect outcome at the inference, affecting the application requirements when employed in 6 G systems. Existing defense mechanisms focus on the attacks in a static nature disregarding the dynamic behavior of attackers. Since the 6 G networks are highly AI-driven, attackers can also learn about the network using AI techniques allowing them to execute more sophisticated dynamic attacks. In this paper, we evaluate how existing defense mechanisms fail in the presence of dynamic poisoning attacks. We also propose novel defense DBFFL against dynamic poisoning attacks in FL, to ensure robustness when deployed in 6 G networks and applications.