FEDXAI-SPLITMED: FEDERATED SPLIT LEARNING WITH DIFFERENTIAL PRIVACY AND EXPLAINABLE AI FOR MEDICAL VISUAL QUESTION ANSWERING
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Medical Visual Question Answering (Med-VQA) faces persistent challenges in protecting patient privacy, explaining model decisions, and reducing communication costs in distributed hospital networks. This paper presents FedXAI-SplitMed, an end-to-end framework that jointly addresses these challenges using Federated Split Learning with a ViT-B/16 encoder split at layer6, client-side MedCGAN augmentation, MPI ring-allreduce gradient compression, and a tri-modal XAI module integrating Grad-CAM++, SHAP DeepExplainer, and cross-modal attention rollout. Only differentially private smashed activations (, ) are transmitted, while upper layer gradients are sparsified using Top-. On VQA-RAD, FedXAI-SplitMed achieves 91.4% closed-ended and 74.8% open-ended accuracy, while reducing communication by 96% (2,100MB to 78 MB) and client computation by 64% compared with vanilla FedAvg. BioSentBERT based evaluation yields 93.1% effective clinical accuracy. A hierarchical aggregation mechanism further consolidates client encoders, with its numerical stability validated independently over 400 synthetic training rounds. The results demonstrate that FedXAI-SplitMed enables privacy preserving, interpretable, and communication efficient distributed Med-VQA.
- Primary area:
- interpretability and explainable AI