TL;DR
This paper demonstrates how a passive eavesdropper, by analyzing HTTP traffic, can reconstruct a significant portion of a user's web browsing history and even link it to real-world identities. The study highlights the vulnerability of users to mass surveillance, even with some privacy tools in place, and quantifies the effectiveness of various mitigation strategies.
What Problem Does This Paper Address?
The paper tackles the growing concern of web tracking and its implications for mass surveillance. While it's known that websites use cookies for tracking, this research investigates the extent to which a passive eavesdropper (like a government agency or an ISP) can leverage this tracking, combined with other readily available information, to build detailed profiles of users. The core problem is quantifying the effectiveness of such surveillance, considering factors like the structure of the web, user browsing behavior, and the adversary's capabilities and limitations (e.g., legal restrictions). Specifically, the authors aim to answer: "How much can an adversary learn about an average user by surveilling web traffic?"
Methods
The authors introduce a novel methodology that combines web measurement and network measurement. Key components include:
- Simulated Browsing: Using a platform called OpenWPM, they simulate realistic user browsing behavior by either visiting random subsets of popular websites (Alexa top 500) or by deriving browsing profiles from anonymized AOL search query logs.
- Network Measurement: They deploy these simulated users on cloud machines in different geographic locations (US, Asia, Europe) and capture all HTTP traffic.
- Adversary Model: They define different adversary scenarios, including unrestricted adversaries and those limited by geographic location (e.g., the "one-end foreign" rule for the NSA) or policy restrictions.
- Cookie Linking Algorithm: A core contribution is an algorithm to detect "identifying cookies" (long-lived, stable, unique, high-entropy values) and then build a "cookie linking graph." This graph connects web pages visited by a user through shared third-party cookies, forming "connected components." The "giant connected component" (GCC) represents the portion of a user's traffic that the adversary can link together.
- Identity Leakage Analysis: They also perform a manual analysis of popular websites to identify instances where personally identifiable information (PII) is leaked in plaintext over HTTP, even after a user logs in.
Datasets
- Alexa Top 500/68 Websites: Used to generate simulated browsing profiles based on website popularity within specific regions.
- AOL Search Query Log Dataset: A large dataset of anonymized search queries used to create more realistic browsing profiles by simulating user searches and visiting the resulting links.
- OpenWPM Measurement Data: The primary dataset generated by the authors' own measurement infrastructure, capturing HTTP traffic from simulated users.
- MaxMind GeoLite2 Country Databases: Used for geolocating IP addresses to determine the physical location of network traffic.
Key Insights and Ideas
- Third-party cookies are powerful linkers: Even if IP addresses change, the presence of multiple third-party cookies on different websites allows an adversary to connect a user's browsing sessions across various sites and times.
- "One-end foreign" rule vulnerability: Users outside the US are particularly vulnerable to US-based surveillance because many popular websites embed third-party resources hosted in the US.
- Plaintext PII leaks are common: A significant percentage of popular websites leak user identities (username, email, etc.) in plaintext over HTTP, even after login, providing a direct link between browsing history and real-world identity.
- Privacy tools offer partial protection: While some tools like Ghostery can significantly reduce the amount of linkable traffic, they cannot eliminate it entirely. Built-in browser settings and other tools like "Do Not Track" have a much smaller impact.
- The web is "flat" for adversaries: The interconnectedness of third-party trackers means that even a small fraction of observed traffic can reveal a large portion of a user's browsing history.
Key Results
- High Clustering Rates: For an unrestricted adversary, the study found that 62-73% of a typical user's browsing history could be reconstructed into a single connected component.
- Non-US User Vulnerability: Non-US users browsing local content still had 59-63% of their traffic linkable by an unrestricted adversary.
- US-Bound Traffic for Non-US Users: When restricted to only observing US-bound traffic, the clustering rate for non-US users dropped significantly but remained surprisingly high (12-20%).
- Identity Leakage: 56% of popular websites with account creation leaked some form of personally identifiable information in plaintext. Of these, 42% leaked unique identifiers like usernames or email addresses.
- Mitigation Effectiveness:
- Blocking all third-party cookies reduced linkable traffic to 30.2%.
- Ghostery was the most effective tool, reducing linkable traffic to 24.2%.
- HTTPS Everywhere offered moderate protection, reducing clustering to 46.1%.
- "Do Not Track" had a negligible effect.
Limitations and Future Work
- Tor Browser Effectiveness: The study acknowledges that using a properly configured Tor browser bundle likely defeats the attack due to its design for "cross-origin identifier unlinkability." However, it notes that preventing PII leakage from websites is not a primary goal of Tor.
- Third-Party Cookie Disabling: While disabling third-party cookies deters the attack, it might not completely stop it due to other stateful tracking mechanisms.
- Logged-in User Frequency: The study performed a manual analysis of identity leaks on popular sites but did not measure how frequently typical users are actually logged into these sites, which could affect the real-world impact.
- Geolocation Accuracy: The study used commercial geolocation data and a custom metric, but acknowledged potential inaccuracies due to incorrect geolocations and network latency.
- Future Work: The authors suggest further research into the prevalence of active third-party content and the effectiveness of other mitigation strategies. They also propose adapting their methodology to study other adversary scenarios, such as those involving active attacks or compelled certificate creation.
Summary updated