An Accessible Library for Machine Learning with Homomorphic Encryption

Abstract

Homomorphic encryption allows computation on encrypted ciphertext so that decrypted plaintext produces a semantically useful computation. Since this comes at a tremendous computation cost for general-purpose computing, the synthesis of these approaches remains in the nascent stages of research and reaches practicability only on small data despite the dire need for privacy-preserving techniques in an age of rampant data surveillance. There is, therefore, a need for an accessible toolkit exposing the power of homomorphic encryption for data scientists to apply to machine learning problems. Privacy-preserving machine learning can have a tremendously positive societal benefit by allowing ethical biomedical research, building public trust about the security of sensitive data, and preventing potential abuses and fraud that result from data breaches in increasingly common hacking incidents. This thesis will attempt to contribute to this direction by building a software library that operates on homomorphically encrypted data but remains approachable to data scientists and simultaneously runs on at least some modest datasets and low-power devices, key limitations that so far have hampered uptake of homomorphic encryption to machine learning. Finally, it will seek to quantify the feasible bounds of data sizes and trade-offs of encryption, decryption, prediction, model training time and the performance penalty incurred from encryption.