TinyFedPrompt | M N Aditya

Deploying medical AI on edge hardware while preserving patient privacy is a hard constraint in real-world clinical settings. TinyFedPrompt solves this by replacing full model fine-tuning with prompt tuning — learning a small set of soft tokens that steer a frozen BioMedCLIP backbone — and distributing the training across clients via federated learning.

Each client trains only its prompt tokens locally; raw patient images never leave the device. The aggregated prompt is then broadcast back, achieving competitive pneumonia detection accuracy at a fraction of the communication and compute cost of standard federated fine-tuning.

Key contributions

First application of federated prompt tuning to a large vision-language medical model (BioMedCLIP).
Runs on Raspberry Pi–class edge hardware; < 5 MB model update per round.
Privacy guarantee: no raw data leaves the client at any point.

Tech stack: Python · PyTorch · CLIP · Flower (flwr) · Raspberry Pi · PACCAS 2026 (Aditya et al., 2026)

References

2026

C
TinyFedPrompt: Federated Prompt Tuning of BioMedCLIP for Privacy-Preserving Pneumonia Detection on Edge Hardware

Moodur Narayanan Aditya, Keerthivasan Krishnamoorthy, Kartikeya Aripirala, and 1 more author

In PACCAS 2026, 2026

Forthcoming

Abs Bib

Deploying medical AI on edge hardware while preserving patient privacy demands both computational efficiency and strong privacy guarantees. TinyFedPrompt replaces full fine-tuning of BioMedCLIP with federated prompt tuning: each client learns a small set of soft prompt tokens locally while the frozen backbone remains on-device. Raw patient images never leave the client, and only compact prompt updates are aggregated each round. The system achieves competitive pneumonia detection accuracy on Raspberry Pi-class hardware with under 5 MB of communication per client per round, demonstrating practical privacy-preserving medical AI at the edge.
@inproceedings{aditya2026tinyfed, title = {{TinyFedPrompt}: Federated Prompt Tuning of BioMedCLIP for Privacy-Preserving Pneumonia Detection on Edge Hardware}, author = {Aditya, Moodur Narayanan and Krishnamoorthy, Keerthivasan and Aripirala, Kartikeya and Jansi, R.}, booktitle = {PACCAS 2026}, year = {2026}, note = {Forthcoming} }