Differential Privacy in the Wild

Abstract

Differential privacy has emerged as an important standard for privacy preserving computation over databases containing sensitive information about individuals. Research on differential privacy spanning a number of research areas, including theory, security, database, networks, machine learning, and statistics, over the last decade has resulted in a variety of privacy preserving algorithms for a number of analysis tasks. Despite maturing research efforts, the adoption of differential privacy by practitioners in industry, academia, or government agencies has so far been rare. Hence, in this tutorial, we will first describe the foundations of differentially private algorithm design that cover the state of the art in private computation on tabular data. In the second half of the tutorial we will highlight real world applications on complex data types, and identify research challenges in applying differential privacy to real world applications.

Membership Inference Attacks on Sequence-to-Sequence Models: Is My Data In Your Machine Translation System?

“…A large body of work exists in differential privacy (Dwork, 2008;Machanavajjhala et al, 2017). Differential privacy provides guarantees that a model trained on some dataset A train will produce statistically similar predictions as a model trained on another dataset which differs in exactly one sample.…”

You Can Use But Cannot Recognize: Preserving Visual Privacy in Deep Neural Networks

“…Luo, Wu, et al [36] presented an approach minimizing trainable parameters, achieving commendable performance in extensive experiments on diverse visual recognition tasks. However, despite the competitive performance of differential privacy in many fields, it faces some challenges in computer vision, particularly in releasing image data [37]. In protecting the privacy of image data, differential privacy needs to add a considerable amount of noise to ensure every pixel in the image is sufficiently protected.…”

Constrained Private Mechanisms for Count Data

“…The model of differential privacy [4], [1], [5] has received a lot of attention in the decade since it was christened, from a variety of communities including systems [6], machine learning and signal processing [7] and data management [8]. For a more thorough overview of the area, there are several detailed surveys [9], [2], [10].…”

“…Experimental Setting. We considered a variety of settings of parameter α (typical values chosen are { 1 2 , 2 3 , 10 11 , 99 100 } and group size n (ranging from 2 up to hundreds).…”