Differentially-Private Orthogonal Tensor Decomposition

Differential privacy has recently received a significant amount of research attention for its robustness against known attacks. Decomposition of tensors has applications in many areas including signal processing, machine learning, computer vision and neuroscience. In this paper, we particularly focus on differentially-private orthogonal decomposition of symmetric tensors that arise in several latent variable models and propose two new algorithms. We compare the performance empirically with a recently proposed algorithm as well as a non-private algorithm. We investigate the performance of these algorithms with varying privacy parameters and database parameters. We show that our proposed algorithms provide very good utility even while preserving strict privacy guarantees.