Collaborative Research: SaTC: CORE: Medium: PRIVIPHY: Physiological Biometric Deidentification through Privacy Preserving Generative AI

This project presents intelligent anonymization methods for preserving the privacy of clients’ bio-signals while retaining data utility for clinical purposes. Bio-signal anonymization methods proposed in this project safeguard clients’ personal data against potential stigmatization, judgment, and discrimination. This fosters patients' participation in healthcare and research studies without fear of identity exposure, thereby enabling the development of efficient data-driven AI models for smart healthcare. This project develops modular and scalable anonymization methods that are suitable for both bio-signals from clinical settings as well as bio-signals acquired from wearable devices in everyday settings. Bio-signal anonymization models proposed in this project are highly adaptable and can be customized for clients across diverse demographics and existing health conditions, achieving a ubiquitous coverage of clients. This project directly impacts the healthcare sector by minimizing regulatory costs, improving trust and confidence between clients and healthcare providers, and delivering high-quality smart healthcare services. This project will also train the next generation of digital healthcare providers in curating clients’ data and developing fundamentally secure smart AI models for healthcare. The first technical thrust develops anonymization models for multi-channel bio-signals through reinforcement learning guided generative deep models. This thrust will design reinforcement learning models to understand critical details of bio-signals for adaptive pruning of anonymization models for different health conditions. This approach balances the competing objectives of obfuscating re-identifiable information while preserving the structural characteristics of bio-signals critical for diagnosis. The anonymization models use conditional and multi-view generative adversarial networks to generate multi-variate bio-signals by sanitizing the original signals. The second technical thrust develops a privacy assessment and evaluation framework for updating anonymization models subject to different attacks. This thrust develops an evaluation framework comprising utility and anonymity metrics to provide feedback on updating bio-signal anonymization models based on utility-anonymity analysis. This enables bio-signal anonymization models to be customizable for clients across diverse demographics and pre-existing health conditions, ensuring both fairness and utility-privacy guarantees. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.