The Fast Mixing Mechanism for Differential Privacy
Two new research papers explore advancements in differential privacy for machine learning and data analysis. The first paper introduces a novel sketching mechanism using fast transforms to improve the runtime of differentially private linear regression, achieving state-of-the-art privacy guarantees. The second paper demonstrates that by introducing correlations in local noise, it's possible to achieve the optimal cost for sum estimation in the local differential privacy model, matching the centralized setting. AI
IMPACT These papers offer new theoretical frameworks for privacy-preserving data analysis, potentially enabling more robust and efficient machine learning applications.