Differential mymargin privacy has been deployed to machine learning platforms to preserve the privacy of data in use. A long neglected but important fact is that data privacy is a non-replenishable resource and should be carefully scheduled to maximize its utility gain. In this work, we propose a new privacy budgeting system—DPlanner, which estimates data blocks’ importance to queries and assigns fractional privacy budget to those data blocks contributing most to a query. The scheduler is novelly designed to include two-fold randomness, which satisfies differential privacy with tight budgets, at the same time guarantees the expected utility in the worst-case query sequence when queries arrive in an online fashion. Experiments in a variety of machine learning settings have shown that our DPlanner outperforms the state-of-the-art schedulers by serving at least 25% more queries, or reducing the total privacy consumption by over 50%.

With the popularity of deep learning applications, the privacy of training data has become a major concern as the data sources may be sensitive. Recent studies have found that deep learning models are vulnerable to privacy attacks, which are able to infer private training data from model parameters. To mitigate such attacks, differential privacy has been proposed to preserve data privacy by adding randomized noise to these models. However, since deep learning models usually consist of a large number of parameters and complicated layered structures, an overwhelming amount of noise is often inserted, which significantly degrades model accuracy. We seek a better tradeoff between model utility and data privacy, by choosing directions of noise w.r.t. the utility subspace. We propose an optimized mechanism for differentially-private stochastic gradient descent, and derive a closed-form solution. The form of the solution makes the mechanism ready to be deployed in real-world deep learning systems. Experimental results on a variety of models, datasets, and privacy settings show that our proposed mechanism achieves higher accuracies at the same privacy guarantee compared to the state-of-the-art methods. Further, we extend the privacy guarantee to a mutual information bound, and propose a general form to the utility-privacy problem.

Private individual information are increasingly exposed through high-dimensional and high-order data, with the wide deployment of learning techniques. These data are typically expressed in form of tensors, but there is no principled way to guarantee privacy for tensor-valued queries. Conventional differential privacy is typically applied to scalar values without a precise definition on the shape of the queried data. Realizing that the conventional mechanisms do not take the data structural information into account, we proposeTensor Variate Gaussian(TVG), a new$(\epsilon,\delta) $-differential privacy mechanism for tensor-valued queries. We further introduce two mechanisms based on TVG with an improved utility by imposing the unimodal differentially-private noise. With the utility space available, the proposed mechanisms can be instantiated with an optimized utility, and the optimization problem has a closed-form solution scalable to large-scale problems. Finally, we experimentally test our mechanisms on a variety of datasets and models, demonstrating that TVG is superior than other state-of-the-art mechanisms on tensor-valued queries.

The wide deployment of machine learning (ML) models and service APIs exposes the sensitive training data to untrusted and unknown parties, such as end-users and corporations. It is important to preserve data privacy in the released ML models. An essential issue with today's privacy-preserving ML platforms is a lack of concern on the tradeoff between data privacy and model utility: a private datablock can only be accessed a finite number of times as each access is privacy-leaking. However, it has never been interrogated whether such privacy leaked in the training brings good utility. We propose a differentially-private access control mechanism on the ML platform to assign datablocks to queries. Each datablock arrives at the platform with a privacy budget, which would be consumed at each query access. We aim to make the most use of the data under the privacy budget constraints. In practice, both datablocks and queries arrive continuously so that each access decision has to be made without knowledge about the future. Hence we propose online algorithms with a worst-case performance guarantee. Experiments on a variety of settings show our privacy budgeting scheme yields high utility on ML platforms.