Robust and Leakage-Resilient Physical-Layer Security for BLE-Based Wireless Body Area Networks under Adversarial and Propagation-Driven Constraints

Wireless Body Area Networks (WBANs) are a critical component of emerging Internet of Healthcare Things (IoHT) applications, enabling real-time monitoring through wearable and implantable sensors. Ensuring the authenticity and security of data from these on-body devices is paramount, especially at the physical layer where decisions must be made before higher-layer cryptographic keys are established. This thesis develops a robust, propagation-based authentication framework for WBANs and investigates the resilience of machine learning classifiers for WBAN signals under adversarial conditions. 

First, we introduce a leakage-resilient radio physical-layer scenario authentication mechanism that reliably distinguishes on-body (legitimate) transmitters from off-body (external or adversarial) transmitters using only the known preamble portion of wireless frames. The framework neutralizes device-specific radio frequency (RF) fingerprints to prevent identity leakage, focusing the decision on propagation characteristics unique to the on-body vs. off-body channels. Evaluated on a multi-device Bluetooth Low Energy (BLE) WBAN dataset with controlled on/off-body scenarios, the proposed scheme achieves near-perfect detection (approx. 100% accuracy and AUC ≈ 1.00) even on devices and sessions never seen in training. It meets strict false-alarm requirements with calibrated thresholds, demonstrating that body-induced creeping-wave propagation effects, rather than hardware signatures, can serve as a dependable security indicator at the physical layer.

Second, we assess the vulnerability of deep learning-based RF classifiers in WBAN settings to adversarial attacks. We develop a convolutional neural network (CNN) model to classify devices and motion contexts from BLE signals, and then subject it to a suite of adversarial perturbations, including gradient-based attacks (FGSM, PGD, and CW) and additive noise. Under benign conditions, the CNN achieves high accuracy (80–90%) on device identification and on-body ix motion recognition tasks. However, we show that small, carefully crafted perturbations can drastically degrade performance: iterative attacks often reduce accuracy to near chance levels. Untargeted attacks that broadly perturb inputs are especially devastating, causing misclassification rates upwards of 95%. These findings expose a serious reliability threat for learning-based WBAN security schemes in adversarial environments. 

The thesis concludes with a synthesis of insights from both studies. By combining a physics-grounded authentication layer with awareness of machine learning vulnerabilities, we outline a multifaceted approach to securing WBANs. Key recommendations include integrating adversarial training, exploring complex-valued neural networks for RF data, and leveraging ensemble methods and randomization to harden WBAN classifiers. Overall, this work contributes a novel lightweight authentication solution for resource-constrained wearable devices and highlights the need for robust machine learning in future body-centric wireless networks.

Committee: Amit Sikder (major professor), Wallapak Tavanapong, and Ashfaq Khokhar