Robust and Energy Efficient Client Selection in Federated Learning using Dynamic Gradient Filtering

Federated Learning (FL) presents a novel methodology that has the potential to improve privacy and security when compared to existing methods. This is achieved through allowing numerous users to collaboratively train a machine learning model without sharing the personal data. Notwithstanding its merits, this paradigm could be exposed to poisoning attacks, where an adversary alters local model updates and inserts them into the aggregation process. There are many methods in the literature to ensure robustness against possible attackers, however, the resource-unreliable nature of clients is often not taken into account. In this paper we propose a Robust and Energy Efficient Client Selection (REECS) algorithms that provide an effective resilience against poisoning attacks. In particular, REECS calculates a score for each local model received and associates it with the residual energy of the corresponding client. This process is used to temporarily eliminate from the aggregation the less performing clients with less available energy resources. The attained results showcase the effectiveness of the solutions in terms of both efficiency and resilience.