EUVIMEDEUVIMED entdecken
Crossref

Quantum Federated Learning for Multimodal Perception: A Proof-of-Concept Vision Study

Atul Rawal

Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 806-813

DOI: 10.1609/aaaiss.v10i1.43493 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Atul Rawal.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

We present a preliminary proof-of-concept study for using quantum federated learning (QFL) with multi-modal sensor data across distributed platforms. Our approach uses a hybrid quantum-classical model in which each sensing modality is first processed by its own classical encoder. The resulting features are then combined using a learned, entropy-regularized gating mechanism, passed through a variational quantum circuit, and trained using federated averaging. Both the classical model parameters and the quantum circuit parameters are shared and aggregated across clients during training. We evaluate the approach in two separate studies using the AU-AIR aerial UAV dataset and the nuScenes autonomous-driving mini split. Each dataset is federated independently across four clients and represented using four sensing modalities. All results are evaluated on held-out test data across five random seeds and compared with an equivalent classical pipeline in which the quantum circuit is replaced by a classical layer of similar size. The results show that the overall QFL pipeline can be trained successfully. Entropy regularization keeps the modality contributions nearly balanced, and the quantum parameters converge under federated aggregation. On the nuScenes categories represented in the mini split, the quantum model achieves a performance of macro averaged F1-score of of 0.869 and an accuracy of 0.913 . However, the classical baseline performs similarly, with the F1 of 0.853, and the difference is not statistically significant. On AU-AIR, the results are weaker as severe class imbalance and limited circuit capacity reduce held-out accuracy to 0.452. Because the overall results are mixed, the study shows that QFL can be applied to real multi-modal sensing data while also highlighting important limitations. The findings emphasize the need for skew-aware metrics, strong classical baselines, and further work on scalable and reliable quantum federated perception.

Zitieren

APA

Rawal, A. (2026). Quantum Federated Learning for Multimodal Perception: A Proof-of-Concept Vision Study. Proceedings of the AAAI Symposium Series, 10(1), 806-813. https://doi.org/10.1609/aaaiss.v10i1.43493

Vancouver

Rawal A. Quantum Federated Learning for Multimodal Perception: A Proof-of-Concept Vision Study. Proceedings of the AAAI Symposium Series. 2026;10(1):806-813. doi:10.1609/aaaiss.v10i1.43493

BibTeX
@article{Rawal2026,
  author = {Rawal, Atul},
  title = {Quantum Federated Learning for Multimodal Perception: A Proof-of-Concept Vision Study},
  journal = {Proceedings of the AAAI Symposium Series},
  year = {2026},
  volume = {10},
  number = {1},
  pages = {806--813},
  doi = {10.1609/aaaiss.v10i1.43493}
}
Kostenlos starten →