Privacy Model: Detect Privacy Leakage for Chinese Browser Extensions

Yufei Zhao, Liqun Yang, Zhoujun Li, Longtao He, Yipeng ZhangIEEE Access 2021 · OpenReview

Abstract

The widespread use of browser extensions has introduced significant privacy leakage risks. While previous research focused on detecting private data transmission, the core challenge lies in determining the "rationality" of such transmissions—whether the data sent is necessary for the extension's function. This paper introduces a "Privacy Model" (PM) framework that establishes baseline permissions for sensitive information and authorized servers for each extension. By analyzing 34,095 extensions from three major Chinese browsers, we identified 2,983 instances of privacy leakage. Our results provide a comprehensive analysis of leakage types, malicious domains, and detection accuracy, offering a robust tool for securing the browser extension ecosystem.

Introduction

Browser extensions offer enhanced functionality but often demand broad permissions, creating a significant privacy surface. Traditional detection methods often flag any transmission of sensitive data as a leak. However, this approach fails to account for legitimate data needs—for instance, a mapping tool requires location data, whereas a simple game does not.

This research addresses the "rationality" problem by focusing on the under-studied Chinese browser market, specifically the QQ and 360 browsers. These platforms present unique challenges due to language barriers and distinct app store ecosystems. The goal is to distinguish between "Reasonable Transmission" and "Abnormal Transmission" by modeling the expected behavior of extensions based on their functional categories.

Methodology

The proposed dynamic detection framework moves beyond passive traffic monitoring to evaluate the intent and necessity of data transmissions.

The Detection Workflow

The framework operates through a multi-stage pipeline:

  1. Extension Classification: Using Natural Language Processing (jieba) and TF-IDF algorithms, extensions are categorized into 15 functional groups (e.g., Shopping Assistant, Crawler, Fun).
  2. Privacy Model (PM) Construction: For each extension, a model is built defining the sensitive data types permitted and the authorized destination servers.
  3. Dynamic Detection: The researchers modified the Chromium source code to capture plaintext network traffic directly from internal browser functions, bypassing protections like HSTS.
  4. Traffic Analysis: A recursive decoder handles Base64 and XOR encryption to uncover hidden URLs or system metadata.
The multi-stage workflow of the extension privacy detection framework, from classification to leakage judgment.

Data and Environment

The study utilized a dataset of 34,095 extensions collected between November 2019 and August 2020. To trigger extension behaviors, the framework visited 10 representative Chinese websites, including Baidu, Taobao, and Weibo, alongside a custom login page designed to detect credential theft.

The Privacy Model and Leakage Judgment

The core of the framework is the interaction between captured network traffic and the established Privacy Model.

The detection framework compares real-time network traffic against the predefined Privacy Model to identify unauthorized data exfiltration.

The system identifies five specific leakage types: History (HL), Location (LL), Browser (BL), OS (OL), and Network address (NL). By mapping these types to functional categories, the system can flag an extension even if it has technical permission to access data, provided that data is irrelevant to its purpose.

Results and Analysis

The framework identified 2,983 leaking extensions. Key findings include:

  • High Precision: The system achieved nearly 100% precision and recall for Location, OS, and Network address leakage. History leakage recall was slightly lower (93-95%) due to sophisticated custom encryption.
  • Category Trends: "Enhancement" extensions were the most frequent leakers by volume (62.6%), while "Web page history" and "Crawler" categories had the highest internal percentage of malicious actors.
  • Fingerprinting: Most leaks involved system metadata (screen resolution, OS version) used for user fingerprinting rather than direct password theft.
  • Malicious Infrastructure: Several domains, such as starwebnet.com and datarating.com, were identified as central hubs for collecting data from hundreds of disparate extensions.

Conclusion

The "Privacy Model" approach effectively bridges the gap between technical detection and behavioral rationality. The study reveals a significant "contaminated library" effect, where developers likely reuse code containing tracking bugs across different ecosystems. Future work will focus on automating rationality judgments using machine learning and advocating for stricter privacy policy mandates in browser extension stores.

Summary updated

Abstract

The wide use of browser extensions brings the privacy leakage problem. The previous works detected private data transmission to find privacy leakage in Chrome or Firefox, but the real challenge is to determine whether the transmission is reasonable because the privacy data that existed in transmission does not absolutely mean leaking. To this end, we establish a privacy model for each extension, which contains the sensitive information permitted to be used and servers authorized to communicate with. In order to evaluate the effectiveness of the proposed method, we develop a dynamic browser extension privacy detection framework. It first builds privacy models for extensions and records all network traffic when accessing test pages. Then, the leakage results are presented according to the strict privacy leakage judgment rules. In this paper, the experiments are conducted in a real environment, and our work is verified by 34,095 extensions which are collected from 3 mainstream browsers in China from November 2019 to August 2020. There is a total of 2,983 extensions that exist privacy leakage. We further conduct a comprehensive analysis of the results including calculating the precision, recall, accuracy, and F1 score for each type of leakage, and show the information leaked by different extension categories and the malicious domain name that collecting the users’ privacy, as well as the results changing of detection over time.
BibTeX
@inproceedings{zhao2021privacy,
  author = {Yufei Zhao and Liqun Yang and Zhoujun Li and Longtao He and Yipeng Zhang},
  title = {{Privacy Model: Detect Privacy Leakage for Chinese Browser Extensions}},
  booktitle = {IEEE Access},
  year = {2021},
  url = {https://openreview.net/forum?id=ChXZDXxE8mR},
}
Figure