PrivacyOracle System Pioneers LLM-Based Privacy Management in Smart Environments
One part of the mHealthHUB Digital Toolbox is a resource of digital health frameworks, including PrivacyOracle: A prototype system for configuring automated privacy decisions within smart-built environments using large language models: mhealthhub.org/portfolio/priv
Modern smart buildings and environments increasingly rely on extensive sensory infrastructures to monitor, optimize, and respond to occupant behaviors. While these sensors enable numerous benefits, they also raise significant privacy challenges. Individuals often lack awareness of the sensors around them, the data being collected, and the ways that information is used, making it nearly impossible to manage privacy decisions manually. This situation is further complicated as the number of sensors and data-driven services continues to grow, leaving occupants with an overwhelming number of privacy choices.
To address this challenge, the mDOT Center’s Srivastava Lab developed PrivacyOracle, a prototype system that leverages large language models (LLMs) to act as an automated privacy firewall. PrivacyOracle is designed to make privacy decisions on behalf of users by reasoning over social and legal norms, understanding sensor contexts, and applying appropriate privacy-preserving transformations when needed. By using LLMs, the system can handle qualitative reasoning tasks that were previously infeasible for automated solutions, such as evaluating social acceptability or regulatory compliance.
Evaluations of PrivacyOracle demonstrate its effectiveness: the system achieved up to 98% accuracy in identifying privacy-sensitive states from sensor data and 75% accuracy in assessing social acceptability of information flows. These results indicate that PrivacyOracle can meaningfully reduce the cognitive and technical burden on occupants, enabling them to maintain control over their personal information without requiring expertise or constant attention.
PrivacyOracle represents a significant step forward in the development of intelligent, user-centric privacy solutions for smart environments. By combining AI-driven reasoning with practical privacy management, it provides a blueprint for future systems that can protect individual privacy in increasingly instrumented spaces while allowing smart environments to operate efficiently and ethically
PrivacyOracle is part of the mDOT Center’s Digital ToolBox
Study citation
Wang, B., Garcia, L. A., & Srivastava, M. (2024). PrivacyOracle: Configuring sensor privacy firewalls with large language models in smart built environments. In Proceedings of the 2024 IEEE Security and Privacy Workshops (pp. 239–245). IEEE. https://doi.org/10.1109/SPW63631.2024.00028