[Paper Review] End-to-End Multi-Tab Website Fingerprinting Attack: A Detection Perspective

Mantun Chen, Yongxin Chen+4 more|arXiv (Cornell University)|Mar 12, 2022

Internet Traffic Analysis and Secure E-voting4 citations

TL;DR

This paper proposes an end-to-end Website Fingerprint Detection (WFD) model that jointly localizes and classifies multiple monitored website traces in untrimmed, multi-tab Tor traffic, eliminating artificial assumptions of single-tab browsing and manual trace trimming. The WFD model, inspired by object detection in computer vision, achieves state-of-the-art performance with mAP up to 80.56% on defended traces and maintains robustness across varying base rates and training data sizes.

ABSTRACT

Website fingerprinting attack (WFA) aims to deanonymize the website a user is visiting through anonymous networks channels (e.g., Tor). Despite of remarkable progress in the past years, most existing methods make implicitly a couple of artificial assumptions that (1) only a single website (i.e., single-tab) is visited each time, and (2) website fingerprinting data are pre-trimmed into a single trace per website manually. In reality, a user often open multiple tabs for multiple websites spontaneously. Indeed, this multi-tab WFA (MT-WFA) setting has been studied in a few recent works, but all of them still fail to fully respect the real-world situations. In particular, the overlapping challenge between website fingerprinting has never been investigated in depth. In this work, we redefine the problem of MT-WFA as detecting multiple monitored traces, given a natural untrimmed traffic data including monitored traces, unmonitored traces, and potentially unconstrained overlapping between them. This eliminates the above assumptions, going beyond the conventional single website fingerprint classification perspective taken by all previous WFA methods. To tackle this realistic MT-WFA problem, we formulate a novel Website Fingerprint Detection (WFD) model capable of detecting accurately the start and end points of all the monitored traces and classifying them jointly, given long, untrimmed raw traffic data. WFD is end-to-end, with the trace localization and website classification integrated in a single unified pipeline. To enable quantitative evaluation in our MT-WFA setting, we introduce new performance metrics. Extensive experiments on several newly constructed benchmarks show that our WFD outperforms the state-of-the-art alternative methods in both accuracy and efficiency by a large margin, even with a very small training set. Code is available at https://github.com/WFDetector/WFDetection

Motivation & Objective

• Address the limitations of existing website fingerprinting attacks that assume single-tab, sequentially visited websites and manually pre-trimmed traces.
• Reframe multi-tab website fingerprinting as a detection problem, where the goal is to locate and classify all monitored traces in raw, untrimmed traffic.
• Develop a unified, end-to-end deep learning model that jointly optimizes trace localization and website classification, avoiding error propagation from separate localization and classification steps.
• Introduce new evaluation metrics—mAP and MBps—for realistic multi-tab website fingerprinting scenarios with overlapping traces.
• Demonstrate robustness and generalization across varying base rates, numbers of single-tab traces per multi-tab trace, and small training data sizes.

Proposed method

• Adapt object detection principles from computer vision to treat each monitored website trace as an object in a long traffic sequence, enabling joint localization and classification.
• Design a unified deep learning architecture with shared feature extraction, scale perception, and two-head prediction (for bounding box regression and class classification) trained end-to-end.
• Introduce a novel compact input representation called 'Burst' to improve model efficiency by reducing input dimensionality while preserving critical traffic patterns.
• Employ a lightweight backbone network and efficient head design to enable high-speed inference (up to 120 MBps) without sacrificing accuracy.
• Train the model on raw, untrimmed traffic data containing monitored, unmonitored, and overlapping traces, directly predicting start/end timestamps and website labels.
• Use mean Average Precision (mAP) and throughput (MBps) as primary evaluation metrics to assess detection accuracy and efficiency in realistic settings.

Experimental results

Research questions

• RQ1 Can a single, end-to-end deep learning model effectively detect multiple overlapping website traces in untrimmed Tor traffic without relying on manual trace trimming or sequential visit assumptions?
• RQ2 How does the proposed WFD model perform compared to state-of-the-art methods in terms of accuracy and efficiency under realistic multi-tab browsing conditions with varying trace overlap and base rates?
• RQ3 To what extent does the WFD model generalize across different numbers of single-tab traces per multi-tab trace and varying training data sizes?
• RQ4 How effective is the WFD model in detecting website traces when defended by the GLUE defense mechanism, particularly FRONT noise?
• RQ5 Does the model maintain high performance when trained on small datasets, reflecting real-world constraints of limited data collection?

Key findings

• The WFD model achieves 80.56% mAP on the DS-19 GLUE dataset under FRONT defense, significantly outperforming the best CDSB series method (15.81% mAP), demonstrating superior robustness against defenses.
• With only 3 single-tab traces per multi-tab trace in training, WFD maintains mAP between 60.55% and 69.33%, showing strong generalization across varying numbers of tabs without retraining.
• Even with small training data, WFD achieves at least 55.46% mAP, compared to 62.02% for normal-sized data, indicating strong data efficiency and practicality in low-data regimes.
• The model’s inference speed reaches up to 120 MBps, and training on small data takes only 64 minutes—significantly faster than baseline methods—highlighting high efficiency.
• Performance improves with increasing numbers of single-tab traces in a multi-tab trace, as the proportion of FRONT-defended traffic decreases, suggesting resilience to defense mechanisms.
• The model remains robust under varying base rates (r), with performance improving as r increases, indicating strong adaptability to real-world browsing patterns.