Differentially Private Datastore Generation for Retrieval-Augmented Inference
Researchers have developed a new framework for creating differentially private datastores, crucial for AI systems that use retrieval-augmented inference. Their hashing-based approach partitions data and adds calibrated noise to ensure individual privacy while maintaining utility. Experiments show a minimal accuracy drop of 2.6% at epsilon=5, and the method significantly reduces the effectiveness of membership inference attacks. AI
IMPACT Enables the secure release of datastores for AI, potentially increasing trust and adoption of on-device AI systems.