EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Privacy Preserving Alzheimer’s Disease Classification Using Multi-Head Attention Transfer via Adaptive Temperature Distillation With Federated Learning

S. Sabari Vasan, P. Jayalakshmi

IEEE Access · 2026 · 14 · 66866-66880

DOI: 10.1109/access.2026.3683070 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von S. Sabari Vasan und 1 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Vellore Institute of Technology, Vellore, Tamil Nadu, India.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The implementation of artificial intelligence in medical imaging, challenges on security of the data and institutional collaboration is major. This research proposes a new research framework that uses a combination of the Multi-Head Attention Transfer (MHAT) and the Adaptive Temperature Distillation (ATD) to carry out privacypreserving classification of Alzheimer’s disease using federated learning, as is the first time in the world. Our model utilizes the nature of heterogeneity in distributed medical imaging data by using temperature-scaled knowledge distillation between heterogeneous client models and a centralized server architecture. The framework does implement a dual-phase training methodology: clients are trained on private datasets with their warm-up training and then iteratively exchanged knowledge through soft label aggregation. The dynamic variation of this distillation intensity (from 5.0 as an initial value with exponential decay) as an adaptive temperature mechanism is implemented to optimize knowledge transfer through 30 rounds of federated rounds. Experimental validation on Multi-class Alzheimer’s disease dataset shows an exceptional performance, which shows 82.25% Accuracy, 82.44% Precision, 82.32% Recall, 82.32% F1-score and a remarkable score of 95.30% for ROC-AUC. The framework has an outstanding communication efficiency with an average of 176MB per round and a computational efficiency of 2.2 seconds per client on average. Fairness analysis shows very small variability (0.09) among the client’s performances, which also means a fair distribution of knowledge among heterogeneous architectures. This work composes a new development in the field of privacy-preserving medical AI, showing that with federated knowledge distillation, centralized modeling performance can be achieved without sharing data and staying strictly for the data location.

Förderung

  • Vellore Institute of Technology, Vellore, Tamil Nadu, India

Zitieren

APA

Vasan, S. S., & Jayalakshmi, P. (2026). Privacy Preserving Alzheimer’s Disease Classification Using Multi-Head Attention Transfer via Adaptive Temperature Distillation With Federated Learning. IEEE Access, 14, 66866-66880. https://doi.org/10.1109/access.2026.3683070

Vancouver

Vasan SS, Jayalakshmi P. Privacy Preserving Alzheimer’s Disease Classification Using Multi-Head Attention Transfer via Adaptive Temperature Distillation With Federated Learning. IEEE Access. 2026;14:66866-66880. doi:10.1109/access.2026.3683070

BibTeX
@article{Vasan2026,
  author = {Vasan, S. Sabari and Jayalakshmi, P.},
  title = {Privacy Preserving Alzheimer’s Disease Classification Using Multi-Head Attention Transfer via Adaptive Temperature Distillation With Federated Learning},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {66866--66880},
  doi = {10.1109/access.2026.3683070}
}
Kostenlos starten →