SynFlowFL: A Dynamic Synaptic Flow Framework for Efficient, Personalized Federated Learning

Federated learning enables distributed machine learning while preserving privacy, yet it faces dual challenges of data heterogeneity and device resource constraints. In this paper, we introduce SynFlowFL, a novel comprehensive framework that uniquely integrates synaptic plasticity theory with dynamic network pruning techniques to address these challenges. SynFlowFL achieves efficient model personalization and extreme compression in federated learning settings. In various image classification benchmarks, SynFlowFL demonstrates superior performance, outperforming existing methods across diverse non-IID data distributions. Notably, SynFlowFL reduces training time by up to 12-fold while maintaining high accuracy, showcasing its significant advantages in handling heterogeneous data and resource-constrained environments. Furthermore, SynFlowFL's hierarchical aggregation and client-specific fine-tuning mechanisms effectively balance global knowledge sharing and local personalization. This work establishes a new paradigm for federated learning system design, potentially revolutionizing the deployment of machine learning models in privacy-sensitive and resource-limited scenarios.