A Systematic Review of Privacy-Preserving Approaches in Deep Learning Applications

Abstract

As machine learning becomes increasingly important in areas that handle sensitive data, preserving data privacy is crucial. We present a comprehensive systematic literature review of privacy-preserving deep learning applications. We categorize research papers in this field from 2017 to May 2024, distinguishing between inference and training, and considering single and multiple data owners. Our taxonomy offers a structured approach to understanding state-of-the-art methods and their suitability for various applications. We further provide real-world examples of each scenario to illustrate their practical relevance and discuss challenges and limitations. In addition to reviewing the existing research literature on privacy-preserving deep learning, we identify gaps in current knowledge and suggest future research directions. Our findings show that execution times are steadily decreasing, making privacy-preserving deep learning increasingly practical. Defending models against privacy attacks still poses challenges and often leads to decreased utility, e.g., in terms of accuracy, one of many trade-offs discussed in this paper. By systematizing a vast body of existing work and providing practical guidelines for technique selection and deployment, our paper serves as a valuable resource for experts, practitioners, and newcomers alike, facilitating cross-disciplinary collaboration and advancements in privacy-preserving deep learning.