Efficient Unconditionally Secure Comparison and Privacy Preserving Machine Learning Classification Protocols.

We propose an efficient unconditionally secure protocol for privacy preserving comparison of $$\\ell $$ℓ-bit integers when both integers are shared between two semi-honest parties. Using our comparison protocol as a building block, we construct two-party generic private machine learning classifiers. In this scenario, one party holds an input while the other holds a model and they wish to classify the input according to the model without revealing their private information to each other. Our constructions are based on the setup assumption that there exists pre-distributed correlated randomness available to the computing parties, the so-called commodity-based model. The protocols are storage and computationally efficient, consisting only of additions and multiplications of integers.