[Paper Review] k-fingerprinting: a Robust Scalable Website Fingerprinting Technique

Jamie Hayes, George Danezis|arXiv (Cornell University)|Sep 2, 2015

Internet Traffic Analysis and Secure E-votingComputer Science17 references94 citations

TL;DR

This paper introduces k-fingerprinting, a robust and scalable website fingerprinting technique using a novel variant of random decision forests to identify web pages—including Tor hidden services—based on encrypted or anonymized traffic. It achieves 85% true positive rate and 0.02% false positive rate in distinguishing 30 monitored hidden services from 100,000 unmonitored pages, outperforming prior methods even under defenses and noisy conditions.

ABSTRACT

Website fingerprinting enables an attacker to infer which web page a client is browsing through encrypted or anonymized network connections. We present a new website fingerprinting technique based on random decision forests and evaluate performance over standard web pages as well as Tor hidden services, on a larger scale than previous works. Our technique, k-fingerprinting, performs better than current state-of-the-art attacks even against website fingerprinting defenses, and we show that it is possible to launch a website fingerprinting attack in the face of a large amount of noisy data. We can correctly determine which of 30 monitored hidden services a client is visiting with 85% true positive rate (TPR), a false positive rate (FPR) as low as 0.02%, from a world size of 100,000 unmonitored web pages. We further show that error rates vary widely between web resources, and thus some patterns of use will be predictably more vulnerable to attack than others.

Motivation & Objective

• To develop a more accurate and scalable website fingerprinting technique that works in realistic open-world settings where clients browse both monitored and unmonitored websites.
• To evaluate the effectiveness of k-fingerprinting against website fingerprinting defenses and in the presence of noisy data, including real-world browsing patterns.
• To analyze which network features (e.g., packet count, timing) leak the most information about web page identity under encryption or anonymization.
• To demonstrate that training on a small fraction of data can yield highly accurate attacks, reducing setup costs for long-term surveillance.
• To show that Tor does not provide meaningful protection against fingerprinting, and that hidden services are distinguishable from standard websites with high accuracy.

Proposed method

• k-fingerprinting employs a modified random decision forest algorithm to extract discriminative features from network traffic traces, enabling classification in open-world settings.
• The method uses a subset of features—particularly packet count and volume—showing they leak more information than complex features like inter-arrival times or ordering.
• It operates in an open-world setting with up to 100,000 unmonitored websites, significantly larger than prior studies (e.g., 5,000 or 25,000 websites).
• The classifier is trained on a small fraction of total data, enabling fast deployment and reducing initial data collection overhead.
• Error rate estimation is performed on the training set to identify and exclude high-misclassification-risk websites, improving overall accuracy.
• The approach is validated across four independent datasets, confirming robustness and generalization across diverse browsing sessions and network conditions.

Experimental results

Research questions

• RQ1 Can a website fingerprinting attack maintain high accuracy in an open-world setting with 100,000 unmonitored websites?
• RQ2 How does k-fingerprinting perform against known defenses such as traffic morphing or decoy pages?
• RQ3 Which network features (e.g., packet count, timing) are most informative for identifying web pages under encryption or anonymization?
• RQ4 To what extent can training on a small fraction of data still yield a highly accurate fingerprinting classifier?
• RQ5 Does browsing via Tor provide meaningful resistance to website fingerprinting, and can hidden services be reliably distinguished from standard websites?

Key findings

• k-fingerprinting achieves a true positive rate of 85% and a false positive rate as low as 0.02% when identifying 30 monitored Tor hidden services among 100,000 unmonitored websites.
• The attack remains effective even when trained on a small fraction of data, demonstrating low setup cost and scalability for long-term surveillance.
• Simple features such as the number of packets in a trace leak more information than complex features like packet inter-arrival times or ordering.
• Error rates vary significantly across websites, and attackers can use training data to identify and exclude high-error targets, improving classifier performance.
• Tor does not provide meaningful protection against fingerprinting: k-fingerprinting achieves high accuracy on both standard websites and Tor hidden services.
• The method outperforms prior state-of-the-art attacks in both accuracy and computational efficiency, with testing time around 0.1 CPU seconds per instance.