Privacy-preserving machine learning outsourcing via lightweight secure arithmetic computation

Researchers propose an actively secure protocol for outsourced secure and private machine learning computation, enabling two parties to perform joint inference with a public model while protecting each party's data privacy. The paper notes that recent similar work mainly targets passive (semi-honest) security and may fail completely against active (malicious) participants who can deviate from the protocol; the work treats neural-network-based secure classification as arithmetic computation over integers and demonstrates protocol efficiency on real arithmetized neural network computations, including a collaborative disease prediction network. The motivating scenario given in the paper is multiple hospitals wishing to share medical records to improve diagnosis and disease prediction but unable to share data in plaintext due to privacy concerns.