SPARK: secure, privacy-aware, and robust algorithms for high-accuracy mobile indoor localization using machine learning

Abstract

Navigation systems are the backbone of modern mobility, essential not only for positioning people and devices but also for enabling intelligent routing, congestion management, and personalized guidance. Outdoor navigation owes much of its success to the Global Positioning System (GPS), which has redefined how we navigate outdoors. However, GPS-based systems falter within indoor environments due to their inability to penetrate walls and obstacles. With society increasingly reliant on indoor positioning or localization—whether in airports, shopping malls, hospitals, smart homes, or factories—this limitation demands innovative alternatives. As a result, the focus has shifted toward wireless technologies such as Bluetooth, ultra-wideband (UWB), and Wi-Fi, which—when coupled with indoor localization techniques like trilateration, triangulation, or fingerprinting—can serve as the building blocks of indoor localization systems. Among these, Wi-Fi-based fingerprinting offers a promising solution, leveraging existing Wi-Fi infrastructure for low-cost and accurate localization. Recent advances in machine learning (ML) further enhance the potential of these systems. Despite these developments, implementing a reliable ML assisted indoor localization system poses several critical challenges, which we aim to address in this dissertation. The ML assisted Wi-Fi fingerprinting-based approach relies on two key phases: (1) an offline phase, where location-tagged Wi-Fi fingerprints are collected and trained on the ML model, and (2) the online phase, where new Wi-Fi fingerprints from user devices are processed through the pre-trained ML model to predict the user's locations in real-time. This approach offers resilience against common environmental disturbances, such as multipath interference and shadowing, while also eliminating the need for a line-of-sight (LoS) between the user and Wi-Fi access points (APs)—a limitation that significantly constrains trilateration and triangulation-based methods—making it particularly robust and practical in real-world indoor environments. Yet, it comes with several shortcomings: (i) Balancing Accuracy, Energy Efficiency, and Latency: Mobile devices have limited CPU power, memory, and battery life, making it challenging to run large ML models. Lightweight ML models must provide quick and reliable predictions while conserving energy and maintaining device performance. (ii) Device Heterogeneity Resilience: Variations in Wi-Fi chipsets, antennas, and firmware noise filtering algorithms across devices cause inconsistencies in RSS measurements, even at the same location. Addressing these differences is key to building reliable models that generalize across diverse hardware. (iii) ML Model Security: Wi-Fi systems are vulnerable to adversarial attacks, which manipulate RSS values and compromise location estimates. Resilience to such attacks is critical to ensure secure indoor localization and enhance the reliability of the indoor localization system. (iv) Long-Term Indoor Localization: Indoor spaces change frequently—furniture moves, crowd densities shift, and AP power levels fluctuate. ML models must maintain reliable performance over time in the face of such environmental changes, without requiring costly or frequent re-calibration. (v) Adaptation to Dynamic Environments: Most ML models are static and fail to maintain accuracy as Wi-Fi RSS signals evolve due to device heterogeneity and long-term environmental changes. Enabling the model to continually adapt to these variations is essential for maintaining high localization accuracy and reliability. (vi) Efficient Data Sourcing and Privacy Preservation: Gathering location-tagged fingerprints is labor-intensive and time-consuming, especially with varying environmental factors. Intelligent data sourcing strategies must minimize manual effort while safeguarding user data through privacy-preserving mechanisms. (vii) Explainability and Transparency of ML Models: Despite delivering high accuracy, ML models often act as black box systems, making it difficult to interpret their predictions. Enhancing model explainability is vital for user trust, regulatory compliance, and debugging localization errors, particularly in dynamic environments. To address these core challenges, this dissertation introduces SPARK, a collection of novel ML frameworks tailored for Wi-Fi fingerprinting-based indoor localization. A fundamental requirement for any model in this domain is the ability to deliver high localization accuracy while remaining energy-efficient enough for deployment on mobile devices. All SPARK frameworks are designed with this constraint in mind. To address the challenge of device heterogeneity, SPARK introduces novel attention-based neural networks and vision transformers to extract device-invariant features from Wi-Fi RSS to enhance robustness across devices with different hardware and software configurations. As indoor environments are constantly evolving, ensuring long-term localization stability is essential. SPARK introduces a novel contrastive learning-based Siamese neural network that enables long-term localization without the need for frequent manual re-calibration of the model. To alleviate the burden of manual data collection, SPARK introduces a novel stacked autoencoder-based data augmentation to generates synthetic fingerprints, thereby minimizing the need for labor-intensive labeling. It also introduces a novel privacy-aware crowdsourcing federated learning method to enable decentralized model training without centralized data storage, thereby preserving user privacy while improving scalability. To ensure the model remains adaptive in ever-changing (dynamic) environments, SPARK introduces novel domain-incremental learning, continual learning, and graph-based learning methods. These methods allow the model to adapt to environmental changes and learn the spatial geometry of indoor layouts, allowing adaptability in dynamic and structurally complex environments. As indoor localization systems grow more distributed and intelligent, security becomes critical. SPARK introduces novel capsule neural networks, curriculum-based adversarial learning, and saliency-driven federated learning methods to defend against cyberattacks, ensuring robustness under adversarial conditions. Finally, SPARK addresses the growing need for explainability in ML systems by introducing a novel logic gate-based method for interpreting ML model decisions. This approach offers transparent decision-making, reveals how traditional ML models behave under noisy conditions, identifies key ML decision influencing features, and enables strategic diagnosis of prediction failures in real-world deployments. Together, the SPARK frameworks address several of the most critical and open challenges in Wi-Fi fingerprinting-based indoor localization. This dissertation guides future research toward the development of domain-specific, scalable, secure, and explainable indoor localization systems, validated through extensive real-world evaluations and analysis.