A Federated Learning Client Selection Method via Multi-Task Deep Reinforcement Learning

Federated Learning (FL) is a privacy-preserving paradigm for training machine learning (ML) models, crucial for data privacy and security protection. It has garnered significant attention from both industry and academia. Typically, clients are selected randomly for training and model aggregation in FL scenarios. However, heterogeneity in data distribution and hardware among devices leads to problems such as slow model convergence, low accuracy, and high computational overhead. To address the issues of statistical heterogeneity and system heterogeneity in FL, this paper proposes an intelligent client selection framework via multitask deep reinforcement learning (DRL). Additionally, two reward functions are introduced to alleviate the heterogeneity problem by maximizing model performance and minimizing system latency. Experimental results on MNIST and CIFAR-10 datasets demonstrate the effectiveness of the proposed method.

I. Introduction

The advancement of artificial i ntelligence has m ade i t particularly important to pursue efficient m achine 1 earning and ensure data privacy. Federated Learning (FL) has emerged as a pivotal paradigm in the field of distributed machine learning, offering a unique approach to building collaborative models while preserving the privacy of individual data sources. Unlike traditional methods reliant on centralized data collection, FL empowers individual devices (e.g., smartphones, IoT [1]) known as edge devices to train models locally with their data. These locally trained models are aggregated to form a global model, avoiding the transmission of raw data to a central server. FL aims to leverage diverse data from edge devices and address concerns about data privacy and security by keeping data localized. It finds applications i n h ealthcare [2], facilitating privacy-preserving predictive modeling with patient data, and in smart homes [3], enabling personalized, context-aware services through sensor data integration.