FedHypeVAE: Federated Learning with Hypernetwork-Generated Conditional VAEs for Differentially-Private Embedding Sharing

Sunny Gupta, Nikita Jangid, Amit Sethi
Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare, PMLR 317:127-134, 2026.

Abstract

Federated learning enables collaborative model development across medical institutions without centralizing sensitive patient data, yet existing embedding-level generative approaches often degrade under non-IID clinical heterogeneity and offer limited formal protection against gradient leakage. We introduce FedHypeVAE, a differentially private, hyper-network based conditional variational framework that generates client-specific decoders and priors from lightweight, trainable client codes. This bi-level formulation personalizes the generative process while ensuring privacy preserving parameter synthesis decoupled from raw medical images. Federated optimization with differential privacy and distributional alignment strategies improves stability and cross-site generalization. The proposed framework unifies personalization, privacy, and domain adaptability within the generative layer, offering a principled solution for privacy-aware representation learning in multi institutional medical imaging. Code: github.com/sunnyinAI/FedHypeVAE

Cite this Paper

BibTeX
@InProceedings{pmlr-v317-gupta26a, title = {FedHypeVAE: Federated Learning with Hypernetwork-Generated Conditional VAEs for Differentially-Private Embedding Sharing}, author = {Gupta, Sunny and Jangid, Nikita and Sethi, Amit}, booktitle = {Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare}, pages = {127--134}, year = {2026}, editor = {Wu, Junde and Pan, Jiazhen and Zhu, Jiayuan and Luo, Luyang and Li, Yitong and Xu, Min and Jin, Yueming and Rueckert, Daniel}, volume = {317}, series = {Proceedings of Machine Learning Research}, month = {20--21 Jan}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v317/main/assets/gupta26a/gupta26a.pdf}, url = {https://proceedings.mlr.press/v317/gupta26a.html}, abstract = {Federated learning enables collaborative model development across medical institutions without centralizing sensitive patient data, yet existing embedding-level generative approaches often degrade under non-IID clinical heterogeneity and offer limited formal protection against gradient leakage. We introduce FedHypeVAE, a differentially private, hyper-network based conditional variational framework that generates client-specific decoders and priors from lightweight, trainable client codes. This bi-level formulation personalizes the generative process while ensuring privacy preserving parameter synthesis decoupled from raw medical images. Federated optimization with differential privacy and distributional alignment strategies improves stability and cross-site generalization. The proposed framework unifies personalization, privacy, and domain adaptability within the generative layer, offering a principled solution for privacy-aware representation learning in multi institutional medical imaging. Code: github.com/sunnyinAI/FedHypeVAE} }
Endnote
%0 Conference Paper %T FedHypeVAE: Federated Learning with Hypernetwork-Generated Conditional VAEs for Differentially-Private Embedding Sharing %A Sunny Gupta %A Nikita Jangid %A Amit Sethi %B Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare %C Proceedings of Machine Learning Research %D 2026 %E Junde Wu %E Jiazhen Pan %E Jiayuan Zhu %E Luyang Luo %E Yitong Li %E Min Xu %E Yueming Jin %E Daniel Rueckert %F pmlr-v317-gupta26a %I PMLR %P 127--134 %U https://proceedings.mlr.press/v317/gupta26a.html %V 317 %X Federated learning enables collaborative model development across medical institutions without centralizing sensitive patient data, yet existing embedding-level generative approaches often degrade under non-IID clinical heterogeneity and offer limited formal protection against gradient leakage. We introduce FedHypeVAE, a differentially private, hyper-network based conditional variational framework that generates client-specific decoders and priors from lightweight, trainable client codes. This bi-level formulation personalizes the generative process while ensuring privacy preserving parameter synthesis decoupled from raw medical images. Federated optimization with differential privacy and distributional alignment strategies improves stability and cross-site generalization. The proposed framework unifies personalization, privacy, and domain adaptability within the generative layer, offering a principled solution for privacy-aware representation learning in multi institutional medical imaging. Code: github.com/sunnyinAI/FedHypeVAE
APA
Gupta, S., Jangid, N. & Sethi, A.. (2026). FedHypeVAE: Federated Learning with Hypernetwork-Generated Conditional VAEs for Differentially-Private Embedding Sharing. Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare, in Proceedings of Machine Learning Research 317:127-134 Available from https://proceedings.mlr.press/v317/gupta26a.html.

Related Material