Scalable BLE Cluster Fingerprinting for High-Density IoT Environments

Bluetooth Low Energy (BLE) is widely used among IoT devices to advertise their presence and exchange data over public channels. Despite security measures such as channel hopping and MAC address randomization, prior work demonstrated that machine learning techniques can leverage BLE packet features to identify individual devices, even in highdensity IoT environments. This thesis examines the feasibility of fingerprinting groups of connected BLE devices, addressing a gap in existing literature that has largely focused on individual device identification. An experimental workflow was designed to evaluate the generalizability of a previously proposed proof-of-concept algorithm developed by [1] across di↵erent device ecosystems. Multiple datasets were collected under varying interference conditions and labeled accordingly, comprising one target device group from either the Samsung or Apple ecosystems. Supervised models trained on Samsung data were evaluated on both Samsung and Apple datasets, revealing limited cross-ecosystem generalization. Retraining the models on a subset of Apple data resulted in a substantial performance improvement. Each ecosystem-specific model was further evaluated on its corresponding high-density dataset, revealing strong performance for Samsung-target classification with the Stochastic Gradient Descent model and near-perfect target identification for Apple devices, albeit with reduced noise discrimination. Among the unsupervised approaches, KMeans was the only clustering algorithm that performed consistently well across both ecosystems. Overall, the results demonstrate that the proposed fingerprinting approach can generalize beyond a single ecosystem when retrained on target-specific data, although further refinement is required to ensure robustness across heterogeneous and dynamic device environments.