Title: Differentially Private Stochastic Gradient Descent for in-RDBMS Analytics

Abstract: This paper studies differential privacy for stochastic gradient descent (SGD) in an in-RDBMS system. While significant progress has been made separately on in-RDBMS SGD and private SGD, none of the major in-RDBMS machine learning frameworks have incorporated differentially private SGD. There are two inter-related issues for this disconnect between research and practice: (1) low model accuracy due to added noise to guarantee privacy, and (2) high development and runtime overhead of the private algorithms. This paper takes a first step to remedy this disconnect and proposes a private SGD algorithm to address both issues in an integrated manner. In contrast to the white-box approach adopted by previous algorithms, we revisit and use the classical technique of output perturbation. While using output perturbation trivially addresses (2), it gives rise to challenges in addressing (1). We address this challenge by providing a novel analysis of the $L_2$-sensitivity of SGD, which allows, under the same privacy guarantees, better convergence of SGD when only a constant number of passes can be made over the data. We then integrate this algorithm, along with the state-of-the-art differentially private SGD, into Bismarck , an in-RDBMS analytics system. Extensive experiments demonstrate that our algorithm can be easily integrated, incurs virtually no overhead, scales well, and most importantly, for a number of datasets yields substantially better (up to 4X) test accuracy than the state-of-the-art algorithms.