PrivWeb: Unobtrusive and Content-aware Privacy Protection For Web Agents

2025-09-15

AI Summary

Current web agents pose user privacy risks due to their reliance on web interface access, yet existing research lacks a systematic, user-centered analysis of these risks. Method: We propose a seamless, content-aware, browser-side privacy protection framework. It employs a lightweight local large language model for real-time UI semantic understanding and sensitive information anonymization; integrates a dynamic privacy classification scheme with a sensitivity-adaptive notification mechanism to preserve user control over high-sensitivity data while minimizing interruptions for low-risk actions; and adopts a non-intrusive interaction design supporting customizable user preferences. Contribution/Results: A user study (N=48) demonstrates that our approach significantly reduces privacy concerns (p<0.01), imposes no additional cognitive load (as measured by NASA-TLX), and improves task satisfaction across diverse scenarios—including travel booking, e-commerce, and information retrieval.

Technology Category

• Humans and AI: Intelligent User Interfaces
• Machine Learning: Privacy
• Application Domains: Humanities & Computational Social Science

Application Category

• Responsible Web: Consent frameworks and practices on the web
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Security and Privacy: Browser and end-point security

Abstract

While web agents gained popularity by automating web interactions, their requirement for interface access introduces significant privacy risks that are understudied, particularly from users' perspective. Through a formative study (N=15), we found users frequently misunderstand agents' data practices, and desired unobtrusive, transparent data management. To achieve this, we designed and implemented PrivWeb, a trusted add-on on web agents that utilizes a localized LLM to anonymize private information on interfaces according to user preferences. It features privacy categorization schema and adaptive notifications that selectively pauses tasks for user control over information collection for highly sensitive information, while offering non-disruptive options for less sensitive information, minimizing human oversight. The user study (N=14) across travel, information retrieval, shopping, and entertainment tasks compared PrivWeb with baselines without notification and without control for private information access, where PrivWeb reduced perceived privacy risks with no associated increase in cognitive effort, and resulted in higher overall satisfaction.

Problem

• Reducing privacy risks from web agents' data access
• Providing unobtrusive user control over sensitive information
• Anonymizing private data on interfaces using localized LLM

Innovation

• Localized LLM anonymizes private information
• Privacy categorization schema with adaptive notifications
• Selective task pausing for sensitive user control