Deep Learning with Gaussian Differential Privacy

Abstract

Deep learning models are often trained on data sets that contain sensitive information such as individuals' shopping transactions, personal contacts, and medical records. An increasingly important line of work therefore has sought to train neural networks subject to privacy constraints that are specified by differential privacy or its divergence-based relaxations. These privacy definitions, however, have weaknesses in handling certain important primitives (composition and subsampling), thereby giving loose or complicated privacy analyses of training neural networks. In this article, we consider a recently proposed privacy definition termed f -differential privacy (Dong, Roth, et al., 2019) for a refined privacy analysis of training neural networks. Leveraging the appealing properties of f -differential privacy in handling composition and subsampling, this article derives analytically tractable expressions for the privacy guarantees of both stochastic gradient descent and Adam used in training deep neural networks, without the need of developing sophisticated techniques as Abadi et al. ( 2016) did. Our results demonstrate that the f -differential privacy framework allows for a new privacy analysis that improves on the prior analysis (Abadi et al., 2016), which in turn suggests tuning certain parameters of neural networks for a better prediction accuracy without violating the privacy budget. These theoretically derived improvements are confirmed by our experiments in a range of tasks in image classification, text classification, and recommender systems. Python code to calculate the privacy cost for these experiments is publicly available in the TensorFlow Privacy library.

GRNN: Generative Regression Neural Network—A Data Leakage Attack for Federated Learning

“…Cryptography based privacy-preserving methods, such as secure aggregation protocol [5], Homomorphic Encryption (HE) [3,27], Diferential Privacy (DP) [7,9,10] have been developed for privacy-preserving learning, for instance, linear regression model [12,35], decision trees [4,6], deep neural networks [19,20,27]. However, the cryptography operations are computationally expensive, and the consistency of visual patterns of images is generally not guaranteed in encrypted data format, which usually leads to a learning model with poor generalization ability.…”

Sebastian Dietz’s Contribution to the Discussion of ‘Gaussian Differential Privacy’ by Dong et al.

“…The red lines are obtained via Corollary 4, while the blue dashed lines are produced by the tensorflow/privacy library. See https://github.com/tenso rflow/ privacy for the details of the setting and more experiments in follow-up work (Bu et al, 2019) [Colour figure can be viewed at wileyonlinelibrary.com] where x * is the unique fixed point of f. We will let the sampling fraction p tend to 0 as T approaches infinity. In the following theorem, a 2 + is a short-hand for (max{a, 0}) 2 .…”

Privacy-Preserving Deep Learning on Machine Learning as a Service—a Comprehensive Survey

“…Another PPDL method that utilizes differential privacy is Bu19 [92]. Bu19 proposes Gaussian Differential Privacy (Gaussian DP) which formalizes the original DP technique as a hypothesis test from the adversaries' perspective.The Table 4 shows the features of our surveyed differential privacy-based PPDL.…”