Skip to content
Conference Open access

MA-FedLoRA: Gated Modality-Aware Federated Adaptation for Heterogeneous Medical Data

Sep 2026 · International Conference on Conceptual Structures · pp. 60-64 · 0 citations · 16 references

Abstract

Foundation models advance multimodal medical AI. However, privacy regulations prevent centralized training on multi-institutional heterogeneous data. Existing Federated Parameter-Efficient Fine-Tuning (FedPEFT) methods directly average local parameters. This ignores semantic differences across modalities and causes negative transfer among heterogeneous medical clients. We propose a Modality-Aware Federated Low-Rank Adaptation (MA-FedLoRA) framework. The framework freezes pre-trained vision-language encoders and applies LoRA for lightweight local adaptation. We introduce a modality-aware affinity aggregation mechanism to assign aggregation weights dynamically based on modality compatibility. We also design a performance-guided adaptive gating strategy to control contributions from unreliable clients. We evaluate the framework on three public Kaggle datasets: chest X-rays, brain CT/MRI scans, and biomedical texts. Results show that MA-FedLoRA transfers knowledge selectively and preserves local adaptation capabilities. Compared to the strongest baseline, MA-FedLoRA observes preliminary gains of 0.8% in worst-client accuracy and 0.5% in worst-client Macro-F1 in a single-seed evaluation. It maintains both overall system utility and worst-client robustness.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.