DUQFL-Prox: Drift-Stable Quantum Federated Learning Boosts Fairness

Quantum federated learning (QFL) allows distributed clients to train quantum neural networks without sharing raw data, critical for privacy-sensitive services like fraud detection and genomic classification. However, heterogeneous client data and noisy quantum optimization cause unstable local updates, client drift, and unfairness. The paper introduces DUQFL-Prox (Deep-Unfolded Quantum Federated Learning with Proximal term), a framework that replaces fixed local optimizers with adaptive unfolded SPSA (Simultaneous Perturbation Stochastic Approximation) updates. Each client also incorporates a proximal term to keep its local model aligned with the global model, while a lightweight controller learns step-specific optimization parameters to improve post-aggregation performance.

Experiments on financial fraud and genomic classification tasks show DUQFL-Prox significantly improves stability, generalization, and client fairness compared to standard QFL baselines. The deep-unfolded approach dynamically adjusts update steps per client, reducing drift even under severe data heterogeneity. The proximal constraint prevents local models from diverging, leading to more consistent global convergence. This work demonstrates that combining deep unfolding with proximal regularization can make quantum federated learning viable for real-world distributed intelligent services where both privacy and fairness matter.