RFLASS: Reinforced Federated Learning with Latent-Space Data Augmentation and Adaptive Sample Selection

Federated Learning (FL) enables collaborative model training across distributed clients while preserving data privacy. However, FL faces significant challenges with Non-IID data, leading to slow convergence and degraded model performance. This paper proposes reinforced federated learning with latent-Space data augmentation and adaptive sample selection (RFLASS), a novel framework to address these issues. RFLASS incorporates two key components: a latent-space data augmentation module and a reinforcement learning-based data selection module. The augmentation module generates synthetic samples aligned with clients' local data distributions, perturbs them with differential privacy noise, and shares them with the server to construct a global dataset. This shared dataset is used to balance local data distributions, mitigating client drift and improving global model generalization. The data selection module utilizes reinforcement learning to dynamically identify and prioritize samples that contribute most to global training objectives, optimizing training efficiency and robustness under heterogeneous data conditions. Experimental results on FEMNIST, CIFAR10, and CIFAR100 demonstrate that RFLASS achieves superior test accuracy and faster convergence compared to state-of-the-art methods, particularly in Non-IID scenarios. These findings highlight the effectiveness of RFLASS in addressing data heterogeneity and improving the overall efficiency of federated learning.