Quantum Secure Probably-Approximately-Correct Learning.

In the relevant context of quantum machine learning, here we address the question, "Can a learner complete the learning securely assisted by a quantum property?" We present a classical-quantum hybrid protocol, which is robust against any malicious attacks on a training dataset. It allows only legitimate learners to perform learning, excluding other intruders. We establish a link between secure learning and its sample-complexity bound in the model of probably-approximately-correct (PAC) learning. Specifically, we show that not only can the lower bound on the learning samples characterize a PAC learner, but also an upper bound can be derived to rule out adversarial learners in our protocol. The security condition here stems from the fundamental quantum no-broadcasting principle; thus, no such condition occurs in the classical regime. Therefore, our protocol realizes an instance of genuinely quantum secure learning. The hybrid architecture of our scheme can offer a practical advantage for implementation in noisy intermediate-scale quantum devices.