A machine-learning data-security technique that guarantees a designated authorized learner achieves superior learning outcomes over an eavesdropping learner by encoding the training labels in quantum states, rather than relying on encryption or access control. Cast in the probably-approximately-correct (PAC) learning framework, it defines a 'learning probability' to quantify performance and derives a guarantee condition that depends only on quantities the authorized learner can measure: the training set size and its noise degree.
In supervised machine learning a learner must produce a hypothesis approximating a target function, which requires sufficient training data; an unauthorized eavesdropping learner who accesses that da…