EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Robust Multimodal Federated Learning Under Incomplete Sensing Modalities for Distributed Sensor Networks

Chia-Hui Liu, Tsung-Ming Hsiao, Shyh-Wei Chen

International Journal of Distributed Sensor Networks · 2026-01 · 2026 (1)

DOI: 10.1155/dsn/6422934 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Distributed Sensor Networks, online erschienen am 06.11.2026. Verfasst von Chia-Hui Liu und 2 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Science and Technology Council.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Distributed sensor networks are increasingly used in healthcare monitoring, mobile sensing, disaster awareness, and edge intelligence applications. However, practical sensor‐network environments often face incomplete sensing modalities, heterogeneous data distributions across nodes, and representation drift between local and global models. These issues make reliable multisensor learning difficult when nodes observe different signal types or lose part of their sensing inputs. To address these challenges, this study proposes distributed multimodal representation learning (DMMRL), a robust multimodal federated learning framework with a coordinated three‐level architecture. At the sensor‐data level, the modality‐aware temporal encoding (MATE) learns modality‐specific representations through attention‐based encoding. At the cross‐modal fusion level, the global modality selective fusion (GMSF) module combines graph‐based interactions with modality‐availability masks to regulate fusion weights. At the cross‐node knowledge level, prototype‐anchored alignment provides shared class‐level references, whereas dual‐level α fusion adaptively integrates global and local knowledge at the feature and parameter levels to support client‐specific adaptation. Experiments use a 20‐client federated simulation across three complementary sensing scenarios: HAR for mobile activity sensing, WESAD for wearable physiological sensing, and CrisisMMD for social multimedia sensing in disaster awareness. In the reference comparison using a Dirichlet concentration parameter of 2.0 and client‐partition seed 42, with approximately 14 % of samples having all modalities available, DMMRL achieves accuracies of 0.9363 on HAR, 0.6370 on WESAD, and 0.7154 on CrisisMMD. Across the five evaluated modality‐completeness settings, DMMRL consistently achieves higher classification accuracy and macro‐averaged F1 (Macro‐F1) scores than the evaluated baselines. These results support DMMRL as an effective framework for multimodal federated learning under the evaluated distributed sensor‐network settings with incomplete sensing modalities.

Förderung

  • National Science and Technology Council

Zitieren

APA

Liu, C., Hsiao, T., & Chen, S. (2026). Robust Multimodal Federated Learning Under Incomplete Sensing Modalities for Distributed Sensor Networks. International Journal of Distributed Sensor Networks, 2026(1). https://doi.org/10.1155/dsn/6422934

Vancouver

Liu C, Hsiao T, Chen S. Robust Multimodal Federated Learning Under Incomplete Sensing Modalities for Distributed Sensor Networks. International Journal of Distributed Sensor Networks. 2026;2026(1). doi:10.1155/dsn/6422934

BibTeX
@article{Liu2026,
  author = {Liu, Chia-Hui and Hsiao, Tsung-Ming and Chen, Shyh-Wei},
  title = {Robust Multimodal Federated Learning Under Incomplete Sensing Modalities for Distributed Sensor Networks},
  journal = {International Journal of Distributed Sensor Networks},
  year = {2026},
  volume = {2026},
  number = {1},
  doi = {10.1155/dsn/6422934}
}
Kostenlos starten →