Data Privacy Protection By Kernel Subspace Projection And Generalized Eigenvalue Decomposition
Konstantinos Diamantaras, TEI of Thessaloniki
Sun-Yuan Kung, Princeton University

Abstract:
Various internet services, including cloud providers and social networks collect large amounts of information that needs to be processed for statistical or other reasons without breaching user privacy. We present a novel approach where privacy protection can be viewed as a data transformation problem. The problem is formulated as a pair of classification tasks, (a) a privacy-insensitive and (b) a privacy-sensitive task. Then privacy protection is the requirement that, given the transformed data, no classification algorithm may perform well on the sensitive task while hurting the performance on the insensitive task as little as possible. To that end, we introduce a novel criterion called Multiclass Discriminant Ratio which is optimized using the generalized eigenvalue decomposition of a pair of between class scatter matrices. We then formulate a nonlinear extension of this approach using the kernel GED method. Our proposed methods are evaluated using the Human Activity Recognition data set. Using the kernel projected data the performance of the User recognition task is reduced by 89% while the Activity recognition task is reduced only by 7.8%.