Asian Journal of Probability and Statistics

This survey reviews research on privacy-preserving mechanisms for machine learning (ML) algorithms. As ML technologies become more pervasive, the need for reliable, secure, and privacy-preserving models becomes critical. Numerous real-world incidents have shown that ML systems can leak private or sensitive information about individuals whose data was used for training or evaluation. Even when raw data is not directly exposed, trained models can be vulnerable to attacks that infer membership, attributes, or other hidden properties about the training data. The present study first outlines the main privacy threats in ML, including de-anonymization and linkage attacks, membership inference, attribute inference/model inversion, model extraction, and property inference attacks. We then review key families of privacy-preserving techniques such as k-anonymity and its variants, differential privacy and its relaxations, federated learning, cryptographic approaches (including secure multi-party computation and homomorphic encryption), and widely used tools and libraries that implement these ideas in practice. For each category, we highlight typical threat models, core ideas, and known limitations and trade-offs with model utility.

Finally, we discuss open challenges and future research directions in privacy-preserving ML, including improving formal guarantees, understanding the impact of composition, defending against increasingly adaptive attacks, enhancing fairness and privacy together, and making privacy-preserving techniques more practical and usable in real-world deployments.