Federated Learning and Convex Hull Enhancement for Privacy Preserving WiFi-Based Device-Free Localization

Device-free localization (DFL) has been acknowledged as one of the most emerging technologies for consumer-electronics applications. To achieve high accuracy, many researchers have developed various neural network models to extract effective features. However, privacy protection has not received enough attentions. In existing DFL systems, user data are uploaded to a central server for processing, which carries the risk of DFL users' data privacy leakage. To fill such this gap, we propose a privacy-preserving DFL approach, named PPDFL. Under a federated learning architecture, PPDFL transfers the trained parameters instead of sending raw data to avoid data privacy leakage. Moreover, we design a convex hull optimization algorithm to optimize the model aggregation mechanism. By calculating convex hull areas of data to assign reasonable weights, the global model to obtain better solutions, which enables high inference accuracy with low data quality. Sufficient experiments based on a real testbed show that PPDFL can achieve 95%-100% accuracy in most cases, and when the data quality is poor, the global model still achieves 80%-90% accuracy, which is improved by 5%-10% compared to the widely used FedAvg algorithm.