CRS-FL: a novel framework for communication efficient, reliable, and scalable decentralized federated learning

Abstract

Federated Learning (FL) enables privacy-preserving distributed training, however existing frameworks suffer from high communication overhead, inefficient client selection, and limited scalability. This research proposes an adaptive FL framework integrating BWO for intelligent client selection and hierarchical DHT-based aggregation to enhance efficiency, robustness, and scalability. BWO optimizes client participation based on local models performance and network conditions, while hierarchical DHT aggregation reduces server dependency and improves fault tolerance. The framework will be implemented using TensorFlow and Keras, tested on MNIST, CIFAR-10, and Fashion-MNIST, and evaluated on communication cost, model accuracy, and scalability. This study advances scalable, adaptive, and communication-efficient FL for large-scale distributed AI applications.