PerFedSAC: Energy-time-aware personalized federated learning via soft actor-critic in resource-constrained IoT

Federated learning (FL) is an emerging distributed training framework that allows multiple edge devices to collaboratively build a global model without uploading raw data, effectively reducing the risk of data leakage. It is particularly suitable for privacy-sensitive Internet of Things (IoT) scenarios. Despite its significant advantages in protecting the data privacy of edge devices, FL still faces several challenges in real-world IoT applications. Factors such as device computational heterogeneity, data distribution diversity, and unstable communication environments make traditional strategies relying on random client selection insufficient to meet the dual requirements of system efficiency and model performance, often leading to unnecessary energy consumption and communication overhead. Moreover, most current methods focus on extracting common features from all clients, neglecting the personalized needs arising from data non independent and identically distributed(IID) and task heterogeneity, which limits the model's convergence speed and generalization performance. To address these issues, we restructured the entire FL system framework and propose a personalized FL soft actor-critic (Per-FL-SAC) algorithm based on the soft actor-critic deep reinforcement learning algorithm. This method introduces a novel evaluation metric, global average local user precision (LUP), to measure the overall performance of FL, and selects appropriate clients in each round to accelerate convergence while controlling time and energy costs. Additionally, to enhance the model's adaptability to data heterogeneity, we integrate a personalized layer structure into the local model to better preserve user-specific features and meet personalized modeling requirements. Extensive experiments were conducted on multiple datasets. Experimental results verify the significant advantages of the proposed Per-FL-SAC algorithm. In various comparative experiments, Per-FL-SAC achieves high convergence speed while effectively controlling time and energy costs, and preserves client-specific information, ensuring the long-term stable operation of the system.