A Review of Research Progress on Utility-Optimized Differential Privacy

Differential privacy, with its rigorous mathematical definition and quantifiable privacy guarantee, has become the de facto standard in the field of data privacy protection. However, the inherent contradiction between the strength of privacy protection and data utility remains a core challenge in its application. How to maximize data utility under a given privacy budget, i.e., the utility optimization problem, is one of the central themes of differential privacy research. Focusing on the theme of utility optimization, this paper systematically reviews utility enhancement techniques under four types of differential privacy models: centralized, local, distributed & shuffle, and hybrid models. It elaborates on the core strategies for utility optimization under each model, including noise calibration and post-processing in the centralized model, improvements in randomized response and coding & dimensionality reduction in the local model, the privacy amplification effect and accuracy leap in the shuffle model, as well as the synergy of hybrid models with techniques such as data pooling, encryption, and anonymization. Based on an in-depth comparative analysis of existing methods, this paper identifies future research challenges and hotspots, including high-dimensional and complex data, adaptive budget allocation, and privacy protection for large models, aiming to provide a systematic reference for privacy-preserving data analysis and further research on differential privacy.