Privacy-preserving machine learning for credit card and healthcare applications: CKKS-based TenSEAL implementation with analysis of speed-security tradeoffs

Software designed to perform both data encryption and Machine Learning (ML), individually, is commonplace. However, an increasingly prevalent class of problems focusing on preserving privacy by performing ML evaluations on encrypted data is underperforming in terms of three key factors: speed, accuracy, and security. In any given application, an optimal balance of these three factors must be found and maintained, a requirement which can be very difficult to fulfill in many circumstances. In this research, we explore the effects of varying degrees of encryption security on speed and accuracy, and we examine whether improving one of these factors hurts one or both of the others. Additionally, we determine how to find the optimal balance of these parameters for our specific application, an idea which can be generalized to the broader problem of ML with Homomorphic Encryption (HE).