Data augmentation-based personalized federated learning for indoor localization with non-IID fingerprint data

Abstract With the growing use of federated learning (FL) in indoor localization, model drift caused by heterogeneous fingerprint data remains a critical challenge. Existing personalized FL methods attempt to mitigate model drift by aligning model parameters, but this often results in high computational costs. In this paper, we propose DA-PFedLoc, a novel data augmentation-based personalized federated learning framework for indoor localization. DA-PFedLoc addresses fingerprint heterogeneity through two key techniques: (i) local fingerprint data augmentation to reduce distribution divergence among clients, and (ii) a contrastive loss that aligns the probability distributions of local and global models, enabling personalized training. Experimental results on two public indoor localization datasets show that DA-PFedLoc achieves a 98.07% reduction in computational overhead relative to the baselines, while simultaneously improving classification accuracy by 4.52%, and also demonstrates faster convergence and improved stability.