Federated Learning for Radar Gesture Recognition Based on Spike Timing-Dependent Plasticity

Radar-based gesture recognition combined machine learning methods can achieve excellent performance and has been utilized in a wide variety of applications. However, large-scale radar signal processing suffers from the risk of privacy leakage and the problem of limited computing resources. In this article, we propose a federated learning (FL) framework based on the spiking neural network (SNN) to solve above limitations in radar gesture recognition. First, we build a distributed FL system with good privacy protection to process radar data collaboratively, which only exchanges the model parameters between clients to train the model, without the need for transferring the data itself. Finally, we train the SNN using the spike timing dependent plasticity with biological interpretability, which is more compatible with the biological characteristics and can significantly reduce the energy consumption. Our creative combination of low-power SNN and FL; thus, address the issue of resource constraint of edge nodes in FL. Furthermore, we introduce the weight pruning technique to the FL training process to reduce the model's communication cost. We conduct experiments on the 8-GHz radar gesture dataset and Google Soli dataset to test the proposed model's validity. We evaluate the model's performance across various participating devices and examine the SNN's capacity to recognize gestures under different data distribution strategies in FL. We also analyze the energy consumption and communication cost of the proposed model, demonstrating the spike timing dependent plasticity-based algorithm is more energy efficient than the traditional backpropagation, as well as the backpropagation through time algorithms.