A0709
Title: Log-Euclidean mechanism for differentially private PCA
Authors: Seung Woo Kwak - Sungkonghoe University (Korea, South) [presenting]
Sungkyu Jung - Seoul National University (Korea, South)
Abstract: Differential privacy (DP) has emerged as a central concept in privacy protection over the past decade. Principal component analysis (PCA) is widely used for dimensionality reduction of complex datasets, but traditional PCA methods can inadvertently disclose sensitive private information. To address this, differentially private PCA algorithms have been developed. DP protects individual privacy by perturbing outputs to account for potential changes caused by variations in input values. The log-Euclidean mechanism is utilized to enhance differentially private PCA and is compared with other private PCA methods based on the Gaussian mechanism, exponential mechanism, and Wishart mechanism. Adding noise in the logarithmic domain allows for reduced noise in the loading vectors, although the eigenvalues of the matrix are not necessarily preserved. The results indicate that the log-Euclidean mechanism provides higher privacy guarantees and requires smaller sample sizes compared to other methods.
