RSSI-Based Fingerprinting of Bluetooth Low Energy Devices

Guillaume Gagnon, Sébastien Gambs, Mathieu Cunche

TL;DR

This paper introduces a novel attack that exploits Received Signal Strength Indication (RSSI) from Bluetooth Low Energy (BLE) advertising packets to re-identify devices, effectively bypassing their address randomization privacy features. The research demonstrates that RSSI patterns can create a unique "fingerprint" for devices, enabling an adversary to track them even when their MAC addresses change.

What Problem Does This Paper Address?

Bluetooth Low Energy (BLE) is ubiquitous in modern devices, from smartphones to fitness trackers. To protect user privacy and prevent tracking, BLE employs address randomization, where devices periodically change their MAC addresses. However, previous research has shown that this mechanism isn't foolproof and can be circumvented using other leaked information like counters or timing. This paper tackles the critical privacy concern of tracking users through their BLE devices by investigating a previously under-explored side-channel: the Received Signal Strength Indication (RSSI). The core problem is that while address randomization aims to make devices unlinkable, there might be other, more subtle characteristics that can be exploited to achieve re-identification.

Methods

The proposed attack leverages the distribution of RSSI values from BLE advertising packets to create a unique "fingerprint" for each device. This fingerprinting approach involves the following key steps:

  1. RSSI Collection: The adversary passively monitors BLE advertising packets and records the RSSI values along with the device's (randomized) address. This is done in two phases: a training phase to build device profiles and a re-identification phase to test the model.
  2. RSSI Profile Generation: For each device, a sequence of RSSI measurements is used to create a profile, represented as a histogram of RSSI values. This histogram is then normalized to form a vector.
  3. Feature Vector Creation: To train a machine learning model, feature vectors are created by concatenating two RSSI profiles. A label is assigned: if the profiles are from the same device, the label is the device's identifier; otherwise, it's a special "non-matched" label.
  4. Linkage Attack (Machine Learning): A machine learning classifier (specifically, HistGradientBoostingClassifier and RandomForestClassifier were evaluated) is trained on these feature vectors. The trained model then takes new feature vectors and predicts whether the two profiles belong to the same device and, if so, identifies which device.

The novelty lies in using RSSI distributions as the primary feature for fingerprinting, rather than relying on packet content or timing.

Datasets

The study utilizes a dataset of RSSI measurements collected by Inria for a BLE-based contact tracing solution. This dataset was gathered from real smartphones during experimental scenarios, including both static and dynamic environments, with participants positioned in various indoor and outdoor locations. The data includes timestamps, sender/receiver identities, coordinates, RSSI values, and device states. The authors note that this dataset is not publicly available.

Key Insights and Ideas

  • RSSI as a Unique Fingerprint: The most significant insight is that the subtle variations in RSSI, influenced by factors like device orientation, antenna characteristics, and environmental reflections, can create a sufficiently unique and stable fingerprint for a BLE device.
  • Bypassing Address Randomization: The attack demonstrates that even with randomized MAC addresses, devices can still be re-identified by analyzing their RSSI "fingerprints." This highlights a fundamental limitation of relying solely on address randomization for privacy.
  • Static vs. Dynamic Environments: The attack is significantly more effective in static environments where devices remain stationary. In dynamic scenarios with mobile targets, the performance drops considerably, suggesting that relative movement and environmental changes can disrupt the RSSI fingerprint.
  • Receiver Multiplicity: The effectiveness of the attack can be boosted by using multiple receivers, allowing for a more comprehensive collection of RSSI data from different vantage points.

Key Results

  • High Accuracy in Static Scenarios: In static scenarios with up to 30 devices, the proposed attack achieved a re-identification accuracy of up to 99%.
  • Learning Algorithm Performance: The HistGradientBoostingClassifier (HGT) consistently outperformed other tested algorithms (RandomForestClassifier and kNN), achieving accuracies above 98% in static scenarios.
  • Impact of Parameters: The accuracy is sensitive to parameters like histogram resolution (nbin) and the number of sequences (nsplit). Higher values generally lead to better performance, with nbin and nsplit around 10-15 showing strong results.
  • Effectiveness of Multiple Receivers: Using multiple receivers significantly improved accuracy, with one scenario achieving a perfect 1.0 accuracy.
  • Reduced Performance in Dynamic Scenarios: In dynamic scenarios where targets and/or the adversary are moving, the attack's performance dropped significantly, with accuracies often below 50%.

Limitations and Future Work

  • Dynamic Scenarios: The attack's effectiveness is considerably reduced in dynamic environments where targets are mobile. The authors suggest that relative movement and environmental changes disrupt the RSSI fingerprint.
  • Countermeasures:
    • Silent Period: Introducing a "silent period" where a device refrains from changing its address until it has significantly moved or the environment has changed could mitigate the attack.
  • Modifying Transmitting Power: Randomly altering the transmission power of BLE devices could add noise to RSSI measurements, making fingerprinting more difficult, though this might impact service quality and energy consumption.
  • Future Work:
    • Adapting to Other Technologies: The researchers plan to adapt this attack to similar wireless technologies like Bluetooth Classic and Wi-Fi.
  • Improving Dynamic Scenario Performance: Investigating methods to enhance the attack's effectiveness in dynamic scenarios, potentially by using multiple receivers or a continuous re-identification strategy.

Summary updated

Abstract

To prevent tracking, the Bluetooth Low Energy protocol integrates privacy mechanisms such as address randomization. However, as highlighted by previous researches address randomization is not a silver bullet and can be circumvented by exploiting other types of information disclosed by the protocol such as counters or timing. In this work, we propose a novel attack to break address randomization in BLE exploiting side information that has not been considered before: Received Signal Strength Indication (RSSI). More precisely, we demonstrate how RSSI measurements, extracted from received BLE advertising packets, can be used to link together the traces emitted by the same device or re-identify it despite address randomization. The proposed attack leverages the distribution of RSSI to create a fingerprint of devices. An empirical evaluation of the attack on various scenarios demonstrate its effectiveness. For instance in the static context, in which devices remain at the same position, th
BibTeX
@inproceedings{gagnon2023rssi,
  author = {Guillaume Gagnon and Sébastien Gambs and Mathieu Cunche},
  title = {{RSSI-Based Fingerprinting of Bluetooth Low Energy Devices}},
  booktitle = {SECRYPT},
  year = {2023},
  doi = {10.5220/0012139600003555},
}
Figure