differentially private covariance estimation
Reviews: Differentially Private Covariance Estimation
Their empirical results show that their algorithm outperforms several other algorithms in the literature in practice. Their theoretical results imply that their algorithm outperforms simple noise addition in the high privacy/small dataset setting. They explicitly consider the sampling procedure for implementing the exponential mechanism in their setting. I appreciated this part of the paper since this is often swept under the rug. The highlight the problems associated to relying on Gibbs sampling and instead use a rejection sampling scheme proposed by Kent et al.