※ 발표시간: 13~14시 As machine learning (ML) has become a universal tool of big data analysis regardless of field, data privacy has emerged one of the most significant issues to be solved for applying ML to real-world applications. Some non-cryptographic methodologies have been applied so far for privacy preservation, but the loss of information is inevitable which leads to significant reduction in data usability. Homomorphic Encryption (HE) has been recognized one of the most appropriate cryptographic primitives for privacy-preserving ML preserving both data privacy and usability, from its beautiful functionality that allows computation over encrypted data without decryption. However, extremely high computational cost of HE computation originated from the large depth of target functions or a number of non-polynomial operations has remained a main bottleneck of applying HE in privacy-preserving ML. In this thesis, we introduce two main methodologies to overcome this limitation. The first one is to modify the existing ML algorithms into HE-friendly form. We instantiate this methodology to logistic regression, which is one of the most popular method for classification, and show the practicality by applying our method to real-world applications including genome-wide association study (GWAS). The second one is to find efficient polynomial approximation of non-polynomial functions, instead of substituting them with some other HE-friendly operations. Based on composite function approximation methods, we develop complexity-optimal HE algorithms for comparison and min/max functions, which are the most frequently used non-polynomial operations in real-world computation.