Seamless Website Fingerprinting in Multiple Environments

Chuxu Song, Zining Fan, Hao Wang, Richard MartinarXiv 2024 · OpenReview

Abstract

Website fingerprinting (WF) attacks identify the websites visited over anonymized connections by analyzing patterns in network traffic flows, such as packet sizes, directions, or interval times using a machine learning classifier. Previous studies showed WF attacks achieve high classification accuracy. However, several issues call into question whether existing WF approaches are realizable in practice and thus motivate a re-exploration. Due to Tor's performance issues and resulting poor browsing experience, the vast majority of users opt for Virtual Private Networking (VPN) despite VPNs weaker privacy protections. Many other past assumptions are increasingly unrealistic as web technology advances. Our work addresses several key limitations of prior art. First, we introduce a new approach that classifies entire websites rather than individual web pages. Site-level classification uses traffic from all site components, including advertisements, multimedia, and single-page applications. Second, our Convolutional Neural Network (CNN) uses only the jitter and size of 500 contiguous packets from any point in a TCP stream, in contrast to prior work requiring heuristics to find page boundaries. Our seamless approach makes eavesdropper attack models realistic. Using traces from a controlled browser, we show our CNN matches observed traffic to a website with over 90% accuracy. We found the training traffic quality is critical as classification accuracy is significantly reduced when the training data lacks variability in network location, performance, and clients' computational capability. We enhanced the base CNN's efficacy using domain adaptation, allowing it to discount irrelevant features, such as network location. Lastly, we evaluate several defensive strategies against seamless WF attacks.

Introduction

Website Fingerprinting (WF) is a side-channel attack where an eavesdropper identifies which website a user is visiting by analyzing patterns in encrypted traffic, such as packet sizes and timing. While prior research has demonstrated high accuracy, it often relies on unrealistic assumptions. Specifically, most models assume attackers can identify the exact start and end of a page load—a task made difficult by modern Single Page Applications (SPAs) and infinite scrolling. Furthermore, previous models often target specific URLs rather than parent domains, making them fragile to content changes, and fail to account for environmental factors like hardware performance and network latency that vary significantly in the real world.

Methodology

To address these gaps, we developed WFNet, a Convolutional Neural Network (CNN) architecture designed for "seamless" classification. Unlike traditional methods that require full page boundaries, WFNet operates on a window of 500 contiguous packets (represented as a 1000-element vector of jitter and size) taken from any point in a TCP stream.

The seamless WF attack model identifies websites from any 500 contiguous packets, bypassing the need for page boundary detection.

Domain Adaptation and Site-Level Classification

A core innovation is the use of Domain Adaptation via a Gradient Reversal Layer. This technique forces the network to learn domain-invariant features, effectively ignoring specific network or hardware characteristics to focus on the website's unique traffic signature. We also shifted the focus to site-level classification, treating entire domains (e.g., Amazon.com) as single classes to capture the aggregate traffic patterns of a site's various sub-pages.

The model architecture incorporates a domain classifier and gradient reversal layer to achieve domain-invariant feature extraction.

Data Collection

We collected a massive dataset of over 1 million web pages across 40 popular websites using Puppeteer. To ensure robustness, data was gathered from eight distinct environments, including university data centers, multiple global cloud providers, and residential home networks. This diversity allowed us to account for hardware variance, ranging from high-performance servers to aged dual-core desktops, which we found significantly impacts packet timing.

Results and Discussion

Our evaluations yielded several key insights:

  • High Accuracy: In a closed-world scenario, WFNet achieved over 90% accuracy across all environments.
  • Environmental Robustness: Domain adaptation significantly improved performance in unfamiliar environments, raising accuracy from under 60% to nearly 80%.
  • The Impact of Hardware: We discovered that a client's CPU power acts as a side-channel; weaker CPUs generate different packet sequences than strong ones, even for the same website.
  • Defense Efficacy: We tested an "Active Injection" defense that inserts dummy packets at specific trigger points. This method reduced attacker accuracy to just 6% with minimal overhead (2% data volume, 6% latency), proving far more efficient than traditional random padding.

Conclusion

Seamless website fingerprinting represents a more realistic and potent threat than previously assumed. By moving away from page-boundary heuristics and accounting for environmental shifts, WFNet demonstrates that eavesdroppers can identify user activity with high precision. However, the success of Active Injection suggests that efficient, low-overhead defenses are possible. Future work will explore open-world scenarios and the impact of newer protocols like QUIC.

Summary updated

Abstract

Website fingerprinting (WF) attacks identify the websites visited over anonymized connections by analyzing patterns in network traffic flows, such as packet sizes, directions, or interval times using a machine learning classifier. Previous studies showed WF attacks achieve high classification accuracy. However, several issues call into question whether existing WF approaches are realizable in practice and thus motivate a re-exploration. Due to Tor's performance issues and resulting poor browsing experience, the vast majority of users opt for Virtual Private Networking (VPN) despite VPNs weaker privacy protections. Many other past assumptions are increasingly unrealistic as web technology advances. Our work addresses several key limitations of prior art. First, we introduce a new approach that classifies entire websites rather than individual web pages. Site-level classification uses traffic from all site components, including advertisements, multimedia, and single-page applications. Second, our Convolutional Neural Network (CNN) uses only the jitter and size of 500 contiguous packets from any point in a TCP stream, in contrast to prior work requiring heuristics to find page boundaries. Our seamless approach makes eavesdropper attack models realistic. Using traces from a controlled browser, we show our CNN matches observed traffic to a website with over 90% accuracy. We found the training traffic quality is critical as classification accuracy is significantly reduced when the training data lacks variability in network location, performance, and clients' computational capability. We enhanced the base CNN's efficacy using domain adaptation, allowing it to discount irrelevant features, such as network location. Lastly, we evaluate several defensive strategies against seamless WF attacks.
BibTeX
@misc{song2024seamless,
  author = {Chuxu Song and Zining Fan and Hao Wang and Richard Martin},
  title = {{Seamless Website Fingerprinting in Multiple Environments}},
  year = {2024},
  eprint = {2407.19365},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.19365},
}
Figure