Differentially-Private Quantum Federated Learning Against Gradient Inversion Attacks

Abstract

Despite the structural advantages for privacy in federated learning (FL), many gradient inversion attacks have widely appeared to endanger the privacy of participants, by exploiting the information during communications. This is still valid for the case where FL is extended to quantum FL (QFL) to utilize quantum advantages. To deal with the privacy leakage during communications, quantum teleportation (QT) can be adopted to QFL for ensuring privacy during communications, by hiding the important information from outsiders. However, it should also address (i) the threats due to privacy attacks from insiders and (ii) the communication efficiency with QT. To tackle these issues, this paper proposes novel QFL adopting quantum local differential privacy (Q-LDP), which is the extension of differential privacy to quantum domains for both communication efficiency and privacy. The proposed algorithm shows privacy and utility enhancement compared to baselines, in terms of reduced attacks and maximized communication efficiency.