NIH Reporter #1DP2LM015910-01

The Dual Role of Memorization: Computational Methods for Balancing Model Performance and Privacy Protection in AI

The project aims to develop novel data science and computational methods to support AI-driven analysis, advancing human health while preserving safeguards and ethical protections by balancing model performance and privacy protection through innovative memorization techniques.

Description

This project proposes an innovative approach to safeguarding privacy in AI by examining the dual role of memorization, quantifying its adverse effects on privacy leakage and contributions to model performance. It characterizes memorization for each training sample, addressing individual differences in privacy risks and utility contributions. The project leverages sample-level risk and influence scores to develop privacy-preserving techniques that optimize the privacy-utility trade-off during training. These techniques will be integrated into standard AI development pipelines, aiding data holders in making informed decisions about deploying emerging models in healthcare and research.

View original record