Mobile_FL: A streamlined FL framework for process optimisation via client clustering using rough c-means algorithm
Akarsh K. Nair, Jayakrushna Sahoo, Linga Reddy Cenkeramaddi, Gaurav Jaswal, Ebin Deni Raj · 2024
Currently, Federated Learning is one of the most widely accepted distributed learning frameworks for privacy-sensitive applications. Despite the popularity gained, FL frameworks struggle to perform well in terms of accuracy and model convergence for scenarios with diverse clients and data distributions. For tackling such issues, the article presents a clustered FL framework, Mobile_FL exploiting model similarities between clients to group them into clusters with collaboratively trainable private datasets. The Mobile_FL framework employs a multi-tier clustering approach with intermediate aggregation ensuring streamlined convergence and model performance compared to existing models. The intermediate servers are termed virtual servers and clustering is performed based on a strategy inspired from the classical rough c means clustering approach. The rough c means ensures computational efficiency with an effective complexity-system performance trade-off compared to complex algorithms. The intermediate layers also act as a deterrent against curious servers, enhancing both system privacy and user anonymity for clients. The framework is also ideally suited for generating personalized FL models for a subgroup of clients, especially in mobile environments with dynamic client populations. The performance of the model is verified through experiments on standard datasets and also subjected to theoretic analysis.