Benchmarking Relaxed Differential Privacy in Private Learning: A Comparative Survey

Differential privacy (DP), a rigorously quantifiable privacy preservation technique, has found widespread application within the domain of machine learning. As DP techniques are implemented in machine learning algorithms, a significant and intricate trade-off between privacy and utility emerges, garnering extensive attention from researchers. In the pursuit of striking a delicate equilibrium between safeguarding sensitive data and optimizing its utility, researchers have introduced various variants of Relaxed Differential Privacy (RDP) definitions. These nuanced formulations, however, exhibit substantial diversity in their underlying principles and interpretations of the core concept of DP, thereby engendering a current void in the comprehensive synthesis of these related works. The principal objective of this article is twofold. Firstly, it aims to provide a comprehensive summary of pertinent research endeavors pertaining to RDP within the realm of machine learning. Secondly, it endeavors to empirically assess the impact on both privacy and utility stemming from machine learning algorithms founded upon these RDP definitions. Additionally, this article undertakes a systematic analysis of the foundational principles underpinning distinct variants of relaxed definitions, culminating in the development of a taxonomy that categorizes these RDP definitions.