Encrypted machine learning of molecular quantum properties

Abstract Large machine learning (ML) models with improved predictions have become widely available in the chemical sciences. Unfortunately, these do not protect privacy necessary within commercial settings, prohibiting use of potentially extremely valuable data by others. Encrypting prediction process can solve this problem double-blind model evaluation and prohibits extraction training or query data. However, contemporary ML based on fully homomorphic encryption federated are either too expensive for practical to trade higher speed weaker security. We implemented secure computationally feasible encrypted using oblivious transfer enabling molecular quantum properties across compound space. we find that kernel ridge regression a million times more than without encryption. This demonstrates dire need compact architecture, including representation matrix size, minimizes costs.