Feature fusion federated learning for privacy-aware indoor localization

In recent years, Indoor Positioning Systems (IPS) have emerged as a critical technology to enable a diverse range of Location-based Services (LBS) across different sectors, such as retail, healthcare, and transportation. Despite their strong demand and importance, existing implementations of IPS face significant challenges concerning accuracy and privacy. The accuracy issue is mainly rooted in the inherent characteristics of Received Signal Strength (RSS), which is widely integrated into current IPS as it only requires readily available WiFi infrastructure. Several studies have demonstrated that RSS suffers from instability and inaccuracy in the presence of environmental changes, making it an inadequate choice for precise IPS. Furthermore, most state-of-the-art IPS encounter privacy and data security issues as they often require users to share their privacy-sensitive location data with a centralized server. Unfortunately, centralized data collection and processing potentially expose users to privacy breaches. To tackle these shortcomings, we advocate for a comprehensive, accurate, and multifaceted solution that enables users to harness the benefits of IPS without provoking privacy concerns. First, we address the positional inaccuracy problem by combining the strengths and synergies between RSS and Channel State Information (CSI). Fusing these complementary metrics delivers increased stability against environmental fluctuations. Thereby, it provides a robust foundation for reliable and accurate positioning outcomes. Second, to address the privacy challenge, we integrate Federated Learning (FL) into the proposed solution to enable the collaborative development of machine learning-based IPS models while ensuring that user data remains decentralized. We conducted a comprehensive assessment to evaluate the performance of the proposed IPS and the corresponding overheads compared to established baseline techniques that utilize either RSS or CSI independently. The results indicate significant enhancements, highlighting our solution's ability to effectively address accuracy and privacy challenges.