A Shapley value-enhanced evaluation technique for effective aggregation in Federated Learning

Abstract

5G networks make it possible to transfer real-time sensory data between millions of devices, forming the internet of things. A typical method to utilize these data is to train a machine learning algorithm to extract the features. Federated learning (FL) is a platform for a coalition of clients to train a model collaboratively without sharing their data to preserve data privacy. Data and model poisoning attacks, free-riding attacks, and model divergence due to clients' non-independent and identically distributed (non-IID) datasets are some challenges in conventional federated learning. The lack of an evaluation method in federated averaging (FedAvg) in FL makes it impossible to identify malicious users or amend the divergence of the global model. In this study, we propose a Shapley-based aggregation algorithm called Shapley averaging (ShapAvg) to aggregate the global model more effectively by evaluating the clients' models. In this algorithm, each client's weight in the weighted average will be proportional to its contribution to the global model performance. The results show that the proposed method outperforms FedAvg when using non-IID datasets and in case of data poisoning or free-riding attacks.