Advances in privacy-preserving machine learning
OA Version
Citation
Abstract
Building useful predictive models often involves learning from personal data. For instance, companies use customer data to target advertisements, online education platforms collect student data to recommend content and improve user engagement, and medical researchers fit diagnostic models to patient data. A recent line of research aims to design learning algorithms that provide rigorous privacy guarantees for user data, in the sense that their outputs---models or predictions---leak as little information as possible about individuals in the training data. The goal of this dissertation is to design private learning algorithms with performance comparable to the best possible non-private ones. We quantify privacy using \emph{differential privacy}, a well-studied privacy notion that limits how much information is leaked about an individual by the output of an algorithm. Training a model using a differentially private algorithm prevents an adversary from confidently determining whether a specific person's data was used for training the model.
We begin by presenting a technique for practical differentially private convex optimization that can leverage any off-the-shelf optimizer as a black box. We also perform an extensive empirical evaluation of the state-of-the-art algorithms on a range of publicly available datasets, as well as in an industry application.
Next, we present a learning algorithm that outputs a private classifier when given black-box access to a non-private learner and a limited amount of unlabeled public data. We prove that the accuracy guarantee of our private algorithm in the PAC model of learning is comparable to that of the underlying non-private learner. Such a guarantee is not possible, in general, without public data.
Lastly, we consider building recommendation systems, which we model using matrix completion. We present the first algorithm for matrix completion with provable user-level privacy and accuracy guarantees. Our algorithm consistently outperforms the state-of-the-art private algorithms on a suite of datasets. Along the way, we give an optimal algorithm for differentially private singular vector computation which leads to significant savings in terms of space and time when operating on sparse matrices. It can also be used for private low-rank approximation.
Description
License
Attribution 4.0 International