A Systematic Review of Challenges and Techniques of Privacy-Preserving Machine Learning

Abstract

Machine learning (ML) techniques are the backbone of Prediction and Recommendation systems, widely used across banking, medicine, and finance domains. ML technique’s effectiveness depends mainly on the amount, distribution, and variety of training data that requires varied participants to contribute data. However, it’s challenging to combine data from multiple sources due to privacy and security concerns, competitive advantages, and data sovereignty. Therefore, ML techniques must preserve privacy when they aggregate, train, and eventually serve inferences. This survey establishes the meaning of privacy in ML, classifies current privacy threats, and describes state-of-the-art mitigation techniques named Privacy-Preserving Machine Learning (PPML) techniques. The paper compares existing PPML techniques based on relevant parameters, thereby presenting gaps in the existing literature and proposing probable future research drifts.