Quantum federated learning through ancilla-driven quantum computation

In the current noisy intermediate-scale quantum era, the limited quantum capabilities of client devices have prompted some quantum federated learning (QFL) schemes to delegate model training tasks to the server with substantial quantum computing resources. While these approaches mitigate the constraints of insufficient quantum resources on the client side, they also introduce challenges such as the requirement for the quantum server to prepare highly entangled brickwork states, clients to prepare quantum initial states, and potential security risks related to the leakage of model parameters and output information. To address these issues, a novel QFL scheme based on ancilla-driven quantum computation is proposed. In this scheme, the quantum server is responsible for preparing qubits and executing non-parameterized entangling gates in the model, while the client performs the parameterized rotation gates, which contains private information by manipulating and measuring a single ancilla qubit sent from the server. Security analysis demonstrates that this scheme can effectively protect the privacy of the client's data, model parameters, and model outputs. Finally, the effectiveness of the proposed scheme is validated through binary classification experiments on the MNIST handwritten digit dataset using the Qiskit platform.