Differential Privacy for Deep and Federated Learning: A Survey

Abstract

Users' privacy is vulnerable at all stages of the deep learning process. Sensitive information of users may be disclosed during data collection, or during training, or even after releasing the trained learning model. Differential privacy (DP) is one of the main approaches proven to ensure strong privacy protection in data analysis. DP protects the users' privacy by adding noise to the original dataset or the learning parameters. Thus, an attacker could not retrieve the sensitive information of an individual involved in the training dataset. In this survey paper, we analyze and present the main ideas based on DP to guarantee users' privacy in deep and federated learning. In addition, we illustrate all types of probability distributions that satisfy the DP mechanism, with their properties and use cases. Furthermore, we bridge the gap in the literature by providing a comprehensive overview of the different variants of DP, highlighting their advantages and limitations. Our study reveals the gap between theory and application, accuracy, and robustness of DP. Finally, we provide several open problems and future research directions.

A Survey of Machine and Deep Learning Methods for Privacy Protection in the Internet of Things

“…Both scenarios adopt differential privacy, homomorphic encryption, and secret-sharing solutions. Finally, El Ouadrhiri et al [ 27 ] review current techniques to protect user privacy in FL. They classify these techniques in three main groups: first, the techniques (k-anonymity, l-diversity, and t-closeness) that protect user privacy before making a dataset available.…”

Untitled

“…The concept of differential privacy was developed as a randomized mechanism to ensure output distribution. The approach focuses on promoting the elements included and removing access to samples engaged in the systems [24]. It enables some focus to prevent access to the type of sample used in the machine learning process.…”

Restrictively Self-sampled and Compressed Local Differential Privacy in Federated Learning

“…Differential privacy was proposed by Dwork et al, which provides an information-theoretic security guarantee that the output of a function is insensitive to any particular record in the dataset [13]. Since it provides a strict, quantifiable, and context-independent means of privacy protection, it has been widely used to enhance data privacy in machine learning (ML).…”

“…However, it has been shown that sharing original local updates damages users' privacy [7,8]. Concern on privacy issues in FL, many privacy-preserved methods, including secure multi-party computing [9], homomorphic encryption [10], differential privacy [11], and other privacy computing technologies, have been widely studied [12][13][14]. Differential privacy (DP) provides better protection because of its powerful theoretical guarantee and low computational overhead [15].…”