Security Access Control for Deep Gaussian Process Regression

This paper proposes a security access control method for deep Gaussian process (DGP) regression. DGPs construct highly expressive models by hierarchically combining multiple Gaussian processes. The proposed method generates scrambled data using random unitary transforms (RUTs) and random permutation matrices (RPMs), enabling DGP regression to be performed in both scrambled input and output domains. This approach enables controlled access to both input data and prediction results. We demonstrate that DGP regression on scrambled data achieves performance equivalent to that on non-scrambled data. Finally, we validate the effectiveness of the proposed method through experiments using clinical diabetes data.