Quantum-Secure Distributed Deep Learning in Optical Networks

Secure multiparty computation allows for the joint evaluation of multivariate functions across distributed users while preserving the privacy of their local inputs. With the rising demand for computationally intensive deep learning inference, the urgency of this field has increased significantly. These computations are often offloaded to cloud servers, exposing vulnerabilities that could compromise clients' data security. To address this challenge, we present a linear algebra engine that harnesses the quantum nature of light for information-theoretically secure multiparty computation using standard telecommunication components. We apply this engine to deep learning, providing rigorous upper bounds on the information leakage of both the deep neural network weights and the client data, utilizing the Holevo and Cramér-Rao bounds. In the MNIST classification task, our approach achieves test accuracies above 96%, with leakage limited to less than 0.1 bits per weight symbol and 0.01 bits per data symbol. This leakage level is an order of magnitude below the minimum bit precision required for accurate deep learning with state-of-the-art quantization techniques. Our results lay the groundwork for practical quantum-secure computation and unlock secure cloud deep learning.