Data Reconstruction Attack Against Principal Component Analysis

Abstract Attacking machine learning models is one of the many ways to measure privacy models. Therefore, studying performance attacks against techniques essential know whether somebody can share information about models, and if shared, how much be shared? In this work, we investigate widely used dimensionality reduction Principal Component Analysis (PCA). We refer a recent paper that shows attack PCA using Membership Inference Attack (MIA). When membership inference PCA, adversary gets access some principal components wants determine particular record was compute those components. assume knows distribution training data, which reasonable useful assumption for attack. With assumption, show make data reconstruction attack, more severe than For protection mechanism, propose guardian first generate synthetic then also compare our proposed approach with Differentially Private (DPPCA). The experimental findings degree successfully attempted recover users’ original data. obtained comparable results DPPCA. number attacker intercepted affects attack’s outcome. work aims answer safe disclose while protecting privacy.