A new method lets devices train personalized quantum neural networks without ever sharing raw model weights.
Researchers propose vFedProtoQNAS, a quantum federated learning approach where each client searches and trains its own quantum neural network architecture suited to its hardware, rather than using one shared design. Instead of averaging parameters across clients - which breaks down when architectures differ - devices share only class-wise prototypes, compact summaries of what each class looks like in a model's latent space. A central server aggregates these prototypes and sends back global versions, which clients use to align their own models. In tests, the approach beat standard federated averaging (FedAvg) by 3.70% accuracy and produced more consistent class representations across devices.
Federated learning's core promise - train collaboratively without centralizing data - has always assumed every participant runs roughly the same model. Quantum hardware breaks that assumption outright: a laptop-simulated QNN and a real quantum processor don't share circuit structure, so adding personalized architecture search just makes the parameter-averaging problem worse. Swapping weight-sharing for prototype-sharing sidesteps the incompatibility rather than solving it, echoing prototype-based federated learning already used for classical neural networks facing similar device mismatches.
The gain here is a 3.70% accuracy bump on small, simulated quantum networks - useful signal, not proof this scales to real quantum hardware.