CAPED: Context-Aware Privacy Exposure Defense for Mobile GUI Agents

The paper introduces CAPED, a context-aware privacy exposure defense layer for mobile GUI agents that selectively masks incidental private content before screenshots are sent to remote multimodal agents. CAPED utilizes task requirements and screen context to minimize privacy leakage, achieving a reduction in success-conditioned weighted seeded leakage from 0.766 to 0.268 while maintaining high task utility. This approach emphasizes the need for task-driven selective exposure, which is critical for enhancing user privacy in mobile applications that rely on visual interfaces.