PAS-QFL: Personalized Ansatz Selection for Quantum Federated Learning
The newly introduced framework, PAS-QFL, tackles the issue of diverse client data in quantum federated learning (QFL). In QFL, various quantum clients work together to train quantum neural networks (QNNs) while keeping their local data private. However, current QFL approaches generally assume a uniform ansatz across clients, neglecting the impact of varying client data on the suitability of the ansatz. With class-imbalanced non-IID data, clients may prefer different ansatz configurations, resulting in inconsistent and biased performance. PAS-QFL separates each client’s QNN into a globally shared ansatz and a personalized private one, customizing the private structure instead of just its parameters. The shared ansatz is prioritized and chosen based on a stability-aware criterion. This framework is detailed in a paper on arXiv (2608.14995), submitted as a cross-abstract, aiming to enhance fairness and stability in QFL by aligning ansatz structures with client data distributions.
Key facts
- PAS-QFL is a Personalized Ansatz Selection framework for QFL.
- It addresses client data heterogeneity in quantum federated learning.
- Existing QFL methods assume a fixed ansatz for all clients.
- Heterogeneous data can lead to unstable and unfair performance.
- PAS-QFL decomposes each client QNN into shared and private ansatz.
- The private ansatz structure is personalized, not just parameters.
- The shared ansatz is selected by a stability-aware cross-client criterion.
- Paper available on arXiv with ID 2608.14995.
Entities
Institutions
- arXiv