Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a hashing-based probability generation framework for creating differentially private datastores that enable secure retrieval-augmented inference in on-device AI systems. By employing locality-sensitive hashing (LSH) to partition high-dimensional data and adding calibrated differential privacy noise, the authors achieve strong privacy guarantees while maintaining minimal accuracy loss. Experimental results demonstrate that at epsilon=5, the proposed method achieves only a 2.6% drop in accuracy and significantly mitigates membership inference attacks, reducing their accuracy to 53.60%.
Achieving strong privacy protection with just a 2.6% drop in accuracy, this method redefines the standards for secure datastore generation in AI applications.
It is crucial for modern on-device AI systems that rely on retrieval-augmented inference to release and share datastores without compromising individual privacy. This can be achieved using Differential Privacy (DP), which provides a formal guarantee that ensures individual contributions remain indistinguishable, even under adversarial analysis. In this paper, we introduce a hashing-based probability generation framework designed to enable the creation and release of differentially private datastores. Our approach employs locality-sensitive hashing (LSH) to efficiently partition high-dimensional data into buckets. We then add calibrated DP noise to the accumulated vote for each bucket, generating a probability distribution across classes. Our method is broadly applicable to any pipeline requiring secure key,value datastore creation and release. We conducted experiments on seven datasets with varying sample sizes and class counts, ranging from 2 to 14. At epsilon=5, our released DP datastore achieves strong privacy protection with only an average 2.6% drop in accuracy. Finally, we benchmark DP datastore resilience to membership inference attacks, reducing attack accuracy to 53.60%.