Approximate homomorphic encryption based privacy-preserving machine learning: a survey

Machine Learning (ML) is rapidly advancing, enabling various applications that improve people's work and daily lives. However, this technical progress brings privacy concerns, leading to the emergence of Privacy-Preserving Machine Learning (PPML) as a popular research topic. In this work, we investigate the privacy protection topic in ML, and showcase the advantages of Homomorphic Encryption (HE) among different privacy-preserving techniques. Additionally, this work presents an introduction of approximate HE, emphasizing its advantages and providing the detail of some representative schemes. Moreover, we systematically review the related works about approximate HE based PPML schemes from the four technical applications and three advanced applications, along with their application scenarios, models and datasets. Finally, we suggest some potential future directions to guide readers in extending the research of PPML.