Federated Learning with Spiking Neural Networks: A Lightweight and Privacy-Preserving Framework for Edge Intelligence

Abstract

Federated Learning (FL) enables privacy-preserving distributed machine learning by keeping raw data local. However, conventional Artificial Neural Networks (ANNs) impose high computational and energy costs. Spiking Neural Networks (SNNs), inspired by biological neurons, provide energy-efficient computation via sparse spike-based processing. This paper presents FL-SNN, a unified framework integrating surrogate gradient-trained SNNs with the FedAvg aggregation protocol. We provide convergence analysis, communication cost evaluation, ablation studies, and comparisons with FedProx and SCAFFOLD.Experiments across MNIST, Fashion-MNIST, CIFAR-10, and asynthetic dataset demonstrate that FL-SNN achieves 94.01 ±0.31% on MNIST, 83.83 ± 0.47% on Fashion-MNIST, 59.6 ± 0.82% on CIFAR-10, and 100% on the synthetic task, while reducing estimated energy by approximately 59%. Index Terms—Federated Learning, Spiking Neural Networks, Edge Computing, Energy Efficiency, Privacy-Preserving Machine Learning, FedAvg, Surrogate Gradient, Convergence Analysis, Neuromorphic Computing