A Novel Fuzzy Inference System-Based Federated Learning Approach

Abstract

Federated learning enables distributed training of machine learning models across a network of decentralised computational nodes, ensuring data privacy by retaining sensitive data on local devices. This paradigm supports the construction of high-performance models by leveraging non-i.i.d. and heterogeneous data dispersed across multiple clients, thereby removing the dependency on centralized data aggregation. Federated learning frameworks utilize a range of model aggregation algorithms—such as FedAvg (federated averaging), FedProx (proximal gradient methods), and FedOpt (adaptive federated optimization)—to achieve convergence by integrating locally updated model parameters into a global model, all while upholding data locality constraints. Fuzzy logic, characterized by its capacity to model vagueness and imprecision through fuzzy sets and linguistic variables, provides a robust mechanism for approximate reasoning in uncertain environments. This study proposes FedFIS, a novel fuzzy logic-based aggregation scheme embedded within the federated learning architecture. The FedFIS methodology circumvents the reliance on gradient-based optimization and computationally intensive mathematical formulations by leveraging fuzzy inference mechanisms, thus offering a computationally lightweight and privacy-preserving alternative for federated parameter aggregation.