Speaker
Description
Federated learning across quantum devices offers a pathway to scalable and privacy-preserving quantum machine learning. However, existing frameworks exchange large gradients or model parameters, creating severe communication and execution overheads that scale with system size and often destabilize training on heterogeneous client data. We introduce Loss Aggregation Quantum Federated Learning (LossAgg-QFL), a framework that transmits only scalar loss values from each client to a central server. Using Simultaneous Perturbation Stochastic Approximation (SPSA), the server infers global parameter updates through constant-cost evaluations, eliminating dependence on qubit number, circuit depth, and parameter dimension. This design reduces per-round communication and circuit executions by up to $10^{6}$ times and improves runtime scaling from $\mathcal{O}(n^{3})$ to $\mathcal{O}(n)$ with respect to qubit count. A multi-directional SPSA strategy further stabilizes optimization, lowering convergence variance from $3.6\%$ to $0.8\%$ across stochastic, gradient-free updates. Evaluations on the MNIST and Fashion-MNIST benchmarks show that LossAgg-QFL achieves up to $+59\%$ and $+33.8\%$ improvements in accuracy under label- and quantity-skew conditions, and also outperforms baseline methods under balanced data. By aligning learning dynamics with quantum-hardware constraints and federated scalability, LossAgg-QFL demonstrates that loss-based aggregation can elevate quantum federated learning from primarily simulation-driven studies to a more practical and resource-efficient paradigm for distributed quantum intelligence.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Quantum-enhanced robustness in machine learning models |