Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Compact Probe Request Fingerprinting with Asymmetric Pairwise Boosting

Giovanni Baccichet, Fabio Palmese, Alessandro E. C. Redondi, Matteo Cesana

TL;DR

This work tackles privacy concerns in Wi‑Fi Probe Requests by proposing a compact, discriminative fingerprint learned from IE fields using Asymmetric Pairwise Boosting. The method searches a large pool of bitmask filters to produce an M-bit binary descriptor, achieving substantial memory reductions while maintaining fingerprinting performance; a 16-bit fingerprint performs close to state-of-the-art with about a 100x reduction in memory. The approach is trained and evaluated on open datasets with 33 devices across 3 channels, and the authors release their code to support reproducibility. The practical impact is enabling scalable device tracking and analytics under tight storage constraints, while highlighting privacy trade-offs inherent in MAC randomization defenses.

Abstract

Probe Requests are Wi-Fi management frames periodically sent by devices during network discovery. Tracking Probe Requests over time offers insights into movement patterns, traffic flows, and behavior trends, which are keys in applications such as urban planning, human mobility analysis, and retail analytics. To protect user privacy, techniques such as MAC address randomization are employed, periodically altering device MAC addresses to limit tracking. However, research has shown that these privacy measures can be circumvented. By analyzing the Information Elements (IE) within the Probe Request body, it is possible to fingerprint devices and track users over time. This paper presents a machine learning-based approach for fingerprinting Wi-Fi Probe Requests in a compact fashion. We utilize Asymmetric Pairwise Boosting to learn discriminating filters which are then used to process specific bit sequences in Probe Request frames, and quantize the results into a compact binary format. Extensive evaluation on public datasets demonstrates a two-order-of-magnitude storage reduction compared to existing methods while maintaining robust fingerprinting performance.

Compact Probe Request Fingerprinting with Asymmetric Pairwise Boosting

TL;DR

This work tackles privacy concerns in Wi‑Fi Probe Requests by proposing a compact, discriminative fingerprint learned from IE fields using Asymmetric Pairwise Boosting. The method searches a large pool of bitmask filters to produce an M-bit binary descriptor, achieving substantial memory reductions while maintaining fingerprinting performance; a 16-bit fingerprint performs close to state-of-the-art with about a 100x reduction in memory. The approach is trained and evaluated on open datasets with 33 devices across 3 channels, and the authors release their code to support reproducibility. The practical impact is enabling scalable device tracking and analytics under tight storage constraints, while highlighting privacy trade-offs inherent in MAC randomization defenses.

Abstract

Probe Requests are Wi-Fi management frames periodically sent by devices during network discovery. Tracking Probe Requests over time offers insights into movement patterns, traffic flows, and behavior trends, which are keys in applications such as urban planning, human mobility analysis, and retail analytics. To protect user privacy, techniques such as MAC address randomization are employed, periodically altering device MAC addresses to limit tracking. However, research has shown that these privacy measures can be circumvented. By analyzing the Information Elements (IE) within the Probe Request body, it is possible to fingerprint devices and track users over time. This paper presents a machine learning-based approach for fingerprinting Wi-Fi Probe Requests in a compact fashion. We utilize Asymmetric Pairwise Boosting to learn discriminating filters which are then used to process specific bit sequences in Probe Request frames, and quantize the results into a compact binary format. Extensive evaluation on public datasets demonstrates a two-order-of-magnitude storage reduction compared to existing methods while maintaining robust fingerprinting performance.

Paper Structure

This paper contains 13 sections, 5 equations, 3 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Candidate bitmask filters
  5. Fingerprint Matching
  6. Dataset
  7. Data Pre-Processing
  8. Experimental Results
  9. Training Data
  10. Selected Filters
  11. Descriptor Length
  12. Clustering Results
  13. Conclusion

Figures (3)

  • Figure 1: Filtering process for a matching pair of . The prediction is also matching: both of the filtered fall on the same side of the threshold.
  • Figure 2: ROC w/o Weighted Hamming Distance
  • Figure 3: Clustering Metrics of the pairwise boosting clustering