I Know What You're Buying: Privacy Breaches on eBay

Tehila Minkus, Keith W. RossPrivacy Enhancing Technologies 2014 · DOI

TL;DR

This paper reveals significant privacy vulnerabilities in eBay's feedback system, demonstrating how attackers can reconstruct a user's purchase history by correlating feedback timestamps and pseudonyms. The authors then extend this to a broad profiling attack, linking eBay usernames to real identities via Facebook, and a category attack, identifying users who bought sensitive items like gun accessories or medical tests.

What Problem Does This Paper Address?

Online marketplaces like eBay rely on reputation systems to build trust in the absence of physical cues. eBay's symmetric feedback system, where both buyers and sellers rate each other, is robust but, as this paper shows, has critical privacy flaws. The core problem is that the public nature of this feedback, combined with specific design choices, allows for the unintended disclosure of users' purchase histories, even for sensitive items, without any action on the buyer's part. This directly contradicts user expectations of privacy for certain types of purchases.

Methods

The paper introduces several attacks that exploit eBay's feedback system:

  • Purchase History Attack: This attack leverages the timestamps associated with feedback entries on both buyer and seller pages. By matching feedback entries with identical timestamps, an attacker can infer which items a buyer purchased from a seller.
  • Pseudonym-based Refinement: To address ambiguity arising from multiple feedback entries sharing the same timestamp, the authors incorporate eBay's pseudonym generation algorithm. Since pseudonyms are generated predictably from usernames, this allows attackers to filter out invalid matches and more accurately reconstruct purchase histories.
  • Broad Profiling Attack: This attack scales the purchase history attack by first obtaining a large list of eBay usernames. These usernames are then cross-referenced with the Facebook Graph API to find potential real-world identities. Once linked, the purchase histories of many users can be reconstructed.
  • Category Attack: This attack focuses on identifying users who have purchased items from specific, sensitive categories. It involves using eBay's search functionality to find sellers of particular items (e.g., gun accessories, medical tests) and then filtering the broad profiling attack results to identify buyers of those items.
  • Side-Information Attack: This attack describes how an attacker with partial knowledge about a target's purchase (e.g., item and date) can use a publicly available database of broad profiles to narrow down potential matches and identify the target's eBay username.

Datasets

  • eBay Feedback Data: The authors collected and analyzed eBay feedback data. Specific datasets mentioned include:
  • A sample database of 5,580 feedback records for analyzing timestamp matching ambiguity.
  • A list of 130,991 eBay usernames for the broad profiling attack.
  • Specific lists of user IDs for case studies: 228,332 for gun accessories, 27,261 for pregnancy tests, and 221 for H.I.V. tests.
  • Facebook Graph API: Used to match eBay usernames with real-world identities.
  • Online Survey Data: A survey of nearly 1,000 eBay users was conducted on Amazon Mechanical Turk to gauge user privacy expectations.

Key Insights and Ideas

  • Timestamp correlation is a powerful, yet overlooked, privacy leak: The seemingly innocuous timestamp on feedback entries can be a direct link to purchase history.
  • eBay's pseudonym system is not as anonymous as it seems: The predictable generation of pseudonyms from usernames provides a crucial heuristic for de-anonymization.
  • The "public by default" nature of feedback has serious implications: Users may not realize that their transaction history, even for sensitive items, can be pieced together without their explicit consent or action.
  • User privacy expectations clash with reality: The survey results indicate that users believe their purchases are more private than they actually are on eBay, especially for sensitive categories.
  • Sensitive purchases are not immune: The paper demonstrates that even purchases of gun accessories and medical tests can be exposed, highlighting the broad applicability of these attacks.

Key Results

  • The purchase history attack, refined with pseudonym filtering, could match 96% of buyers' feedback records to a single seller feedback record with purchase details.
  • The broad profiling attack successfully matched 17.2% of eBay usernames to potential real names on Facebook.
  • For matched users in the broad profiling attack, an average of 18 purchases per person were recovered, complete with descriptions.
  • Case Studies:
    • Gun Accessories: Identified 35,262 suspected gun owners from 228,332 gun accessory purchases.
  • Medical Tests: Identified 4,694 users who purchased fertility-related tests and 37 users who purchased H.I.V. tests.
  • Survey Results:
    • 10.7% of users reported buying health-related items and 4.3% bought sexual/adult items.
  • eBay was chosen over physical stores by 38.6% of users for purchasing sensitive or embarrassing items.
  • A significant majority (76.6%) understood that feedback was public, but only 8.7% correctly identified that "anyone, even not signed in" could see their purchases.

Limitations and Future Work

  • False Positives: The paper acknowledges that the username-to-Facebook matching has a potential for false positives, especially for common names. The 17.2% match rate is considered an upper bound.
  • Data Availability: The effectiveness of some attacks relies on the availability of public eBay data and Facebook's API. Changes to these platforms could impact the attacks.
  • User Awareness: While the survey highlights a disconnect between user expectations and reality, further work could explore more effective ways to educate users about these privacy risks.
  • Recommendations: The authors propose several recommendations, including making private listings the default, using non-persistent pseudonyms for buyers, and encouraging users to avoid username reuse across different platforms. Future work could involve evaluating the effectiveness of these proposed changes.

Summary updated

BibTeX
@inproceedings{minkus2014i,
  author = {Tehila Minkus and Keith W. Ross},
  title = {{I Know What You're Buying: Privacy Breaches on eBay}},
  booktitle = {Privacy Enhancing Technologies},
  year = {2014},
  doi = {10.1007/978-3-319-08506-7\_9},
}
Figure