SHIELD: A Framework for Efficient and Secure Machine Learning Classification in Constrained Environments

Machine learning classification has enabled many innovative services, e.g., in medicine, biometrics, and finance. Current practices of sharing sensitive input data or classification models, however, causes privacy concerns among the users and business risk among the providers. In this work, we resolve the conflict between privacy and business interests using Secure Two-Party Computation. Concretely, we propose SHIELD, a framework for efficient, and accurate machine learning classification with security in the semi-honest model. Building on SHIELD, we realize several widely used classifiers and real-world use cases that compare favorably against related work. Departing definitively from prior works, all of SHIELD's protocols are designed from the ground up to enable secure outsourcing to untrusted computation clouds enabling even constrained devices to handle our most complex use cases in (milli)seconds.