Privacy-Preserving Machine Learning Based on Cryptography: A Survey

Machine learning has profoundly influenced various aspects of our lives. However, privacy breaches have caused significant unease and concern among the general public. Preserving the privacy of sensitive data during the training and inference phases of machine learning is a key challenge. Cryptography-based privacy-preserving machine learning (crypto-based PPML) offers a viable solution to this challenge. In this article, we studied over 100 publications on crypto-based PPML frameworks published between 2016 and 2024, including 55 client-server architecture frameworks and 64 multi-party architecture frameworks. We provide a comprehensive overview of these frameworks, highlighting their features across various dimensions. Furthermore, we conduct an in-depth analysis, delving into scenarios, privacy goals, threat models, and optimization techniques that underpin these innovative solutions. We also discuss the challenges in the field of crypto-based PPML, including aspects of security and privacy, efficiency, and availability and usability. Finally, we offer an outlook on future research directions, aiming to provide valuable insights for both scholars and practitioners.