AdaFed-LDR: Adaptive Federated Learning with Layerwise Dynamics Regularization for Robust Wi-Fi Localization

Wi-Fi Channel State Information (CSI)-based indoor localization enables high-precision positioning, but its deployment across multiple environments faces two major challenges: privacy concerns from centralizing CSI data, and severe statistical heterogeneity (non-IID) arising from the strong environment-dependency of CSI. This heterogeneity creates a stability-plasticity trade-off in federated learning-maintaining precision in known environments (stability) while adapting to unseen domains (plasticity). To address this trade-off, we propose AdaFed-LDR, which combines server-side Confidence-Weighted Adaptive Aggregation with client-side Layerwise Dynamics Regularization (LDR). The aggregation recalibrates client contributions based on feature covariance changes, while LDR imposes depth-dependent constraints-stronger constraints on shallow layers to preserve environment-agnostic features and weaker constraints on deeper layers to allow environment-specific adaptation. Evaluated across 8 indoor environments using Leave-One-Out Cross-Validation and 5 random seeds, AdaFed-LDR achieved a mean localization error (MLE) of 0.41 cm in known environments, corresponding to an 88.2% reduction compared with FedAvg. In domain generalization to unseen environments, AdaFed-LDR achieved an MLE of 218.2±2.8 cm, demonstrating an improvement over FedPos (257.6±14.04 cm). With one adaptation sample per reference point, MLE improved to 21 cm. Ablation experiments confirmed that combining the two proposed components achieved the highest improvement (83.9%) compared with applying them individually, supporting AdaFed-LDR as a reproducible approach to the stability-plasticity trade-off in federated CSI-based localization.