VarFed: Developing Distributed Learning Frameworks for Streaming Data

Federated learning (FL) stands as a pivotal innovation in the realm of machine learning, championing data privacy and decentralised learning in the face of burgeoning data volumes and privacy concerns. Despite its potential, FL's efficacy is often hampered by the challenges posed by non-IID data distributions and the dynamic nature of data environments. This paper presents VarFed, a novel federated learning protocol designed to dynamically adjust aggregation weights in response to streaming data, offering a nuanced approach to model aggregation that transcends the limitations of traditional methods like FedAvg. Central to VarFed is the introduction of Aggregation Weight neural Networks (AWN), which trains and predicts aggregation dynamically based on the local models submitted by the clients. Through extensive experiments, VarFed demonstrably outperforms the conventional FL approach, FedAvg and a stacking-based ensemble learning method, SEL. These experimental results show VarFed's superior adaptability to non-IID and dynamic data.