Abstract
Remote deep learning paradigm raises important privacy concerns related to clients sensitive data and deep learning models. However, dealing with such concerns may come at the expense of more client-side overhead, which does not fit applications relying on constrained environments. In this paper, we propose a privacy-preserving solution for deep-learning-based inference, which ensures effectiveness and privacy, while meeting efficiency requirements of constrained client-side environments. The solution adopts the non-colluding two-server architecture, which prevents accuracy loss as it avoids using approximation of activation functions, and copes with constrained client-side due to low overhead cost. The solution also ensures privacy by leveraging two reversible perturbation techniques in combination with paillier homomorphic encryption scheme. Client-side overhead evaluation compared to the conventional homomorphic encryption approach, achieves up to more than two thousands times improvement in terms of execution time, and up to more than thirty times improvement in terms of the transmitted data size.