Abstract
Website fingerprinting (WF) is a well-known threat to users' web privacy. New internet standards, such as QUIC, include padding to support defenses against WF. Previous work only analyzes the effectiveness of defenses when users are behind a VPN. Yet, this is not how most users browse the Internet. In this paper, we provide a comprehensive evaluation of QUIC-padding-based defenses against WF when users directly browse the web. We confirm previous claims that network-layer padding cannot provide good protection against powerful adversaries capable of observing all traffic traces. We further demonstrate that such padding is ineffective even against adversaries with constraints on traffic visibility and processing power. At the application layer, we show that defenses need to be deployed by both first and third parties, and that they can only thwart traffic analysis in limited situations. We identify challenges to deploy effective WF defenses and provide recommendations to address them.
Introduction
Website Fingerprinting (WF) is a privacy-shattering technique where an eavesdropper—such as an ISP or a malicious Wi-Fi operator—infers which website a user is visiting by analyzing the "shape" of encrypted traffic, including packet sizes, timing, and direction. While the new QUIC transport protocol includes a PADDING frame specifically designed to combat this threat, most existing research assumes users are behind a VPN. In a VPN scenario, the attacker only sees a single encrypted tunnel. However, most users browse the web directly, allowing attackers to see destination IP addresses for various resources like images or scripts. This study evaluates whether QUIC’s padding actually protects the average user in this realistic "open web" setting.
Methods and Adversary Models
The researchers evaluated WF attacks using both traditional machine learning (Random Forest) and deep learning (Var-CNN). They considered two types of adversaries:
- Unconstrained: A powerful attacker with visibility into all traffic and unlimited computational resources.
- Constrained: A more realistic attacker with limited network visibility (observing only specific hops) or limited processing power (analyzing sampled data like NetFlow).
The study simulated several defense strategies, ranging from network-layer defenses (hiding packet sizes and timings) to application-aware defenses where padding is informed by the specific resources being loaded.
Datasets
To ensure a robust evaluation, the authors collected several specialized datasets using Firefox and Chromium browsers. These datasets cover various scenarios, including landing pages, temporal stability, and heterogeneous traffic involving subpages.
The datasets include:
- MAIN & LARGE: 150 and 350 landing pages from sites behind a major CDN.
- TIME: Traces collected six months apart to test fingerprint stability.
- HET (Heterogeneous): Traces including subpages to move beyond simplified "landing page only" models.
- Browser-specific: Data comparing fingerprints between Firefox and Chromium.
Key Insights
The "Global Info" Trap
Hiding the size of individual packets is largely ineffective if an attacker can still observe the total volume of data transferred. Most network-layer defenses fail because they cannot successfully obscure the aggregate size of a website's traffic.
Google as a Universal Fingerprint
Many websites rely on third-party resources from giants like Google (e.g., fonts, analytics). Attackers can identify a "parent" website with approximately 77% accuracy simply by timing the ClientHello packets sent to Google IPs. This method requires significantly less data than observing the full traffic trace.
The Coordination Challenge
Effective application-layer defense requires coordination between the primary website and all third-party trackers or CDNs. If even one third-party provider fails to implement padding, the entire site remains vulnerable to fingerprinting.
Key Results
- Ineffectiveness of Padding: On its own, network-layer padding only reduced attacker accuracy (F1-score) by about 3.5%.
- High Overhead for Dummy Traffic: To significantly reduce attacker accuracy to ~70%, a site would need to inject dummy packets resulting in a prohibitive 231% bandwidth overhead.
- Vulnerability of Subpages: Even if an attacker has not seen a specific subpage, they can often identify the domain because subpages share many third-party resources with the landing page.
- Resilience to Sampling: Even when attackers only see 1% of packets (simulating high-speed monitoring), they achieve a 41.7% F1-score, which is significantly better than random guessing.
Conclusion and Future Directions
The study concludes that current QUIC padding implementations are insufficient for protecting users on the open web. Future work should focus on standardizing resource sizes into fixed-size buckets, utilizing QUIC features like MASQUE to obscure destination IPs, and developing standards through bodies like the W3C to allow first and third parties to coordinate defensive traffic patterns.
Summary updated