Task-Specific Federated Learning for Classification and Regression in WiFi Fingerprinting Indoor Localization: Addressing Non-IID Data and Client Dropouts

With the growing demand for indoor positioning, deep learning-based WiFi fingerprint positioning technology has attracted increasing attention. To address challenges such as the high cost of data collection and privacy risks, Federated Learning (FL) has been introduced into indoor positioning in recent years. However, the practical application of FL is often hindered by performance degradation caused by non-independent and identically distributed (Non-IID) data and client dropouts (straggler). This paper proposes a task-specific WiFi fingerprint localization algorithm based on federated learning to mitigate the performance degradation under Non-IID data and client dropouts. For the classification task of building-floor localization, an improved autoencoder classification model is introduced. The proposed classification model further extracts fingerprint features using a Multi-Layer Perceptron (MLP) enhanced with residual connections. Federated Prototype Learning (FedProto) is employed as the aggregation algorithm, with optimizations applied to the prototype aggregation formula. For the regression task of precise location estimation, an enhanced Vision Transformer (ViT) model is proposed. This model deeply integrates the principles of Support Vector Regression (SVR). The loss function is redesigned to achieve joint optimization from feature extraction to coordinate regression. An adaptive aggregation federated learning algorithm is used for aggregation, which involves a rational stratification of parameters to enhance the model's capability in handling heterogeneous data. Experiments conducted on Non-IID datasets show that the classification model achieved a high accuracy of 99.47% with a relatively small number of parameters. The regression model maintained a low average error distance of 4.35 meters, demonstrating high accuracy even in scenarios with client dropouts. The experimental results fully validate the effectiveness of the proposed algorithm.