FedMal-XAI: an explainable federated vision transformer leveraging knowledge distillation for privacy-preserving malaria detection

Plasmodium parasites are the cause of malaria, a deadly illness that continues to pose a serious danger to world health, especially in areas with low resources where subjectivity, complexity, along with privacy issues make it difficult to employ traditional diagnostic techniques like microscopy and quick diagnostic testing. To overcome these specific challenges of diagnostic subjectivity, logistical complexity, and data privacy, this paper suggests a privacy-preserving federated learning system that uses sophisticated Vision Transformers (ViTs) for automated malaria identification from blood smear images. This paper suggests a privacy-preserving federated learning system that uses sophisticated Vision Transformers (ViTs) for automated malaria identification from segmented red blood cell (RBC) images in order to get around these issues. This architecture successfully addresses important privacy and logistical restrictions by enabling cooperative training among decentralized institutions without exchanging sensitive data. Prominent centralized convolutional neural networks (CNNs) are matched in diagnostic accuracy by the federated ViT models, which include ViT-B/16, DeiT-Tiny, Swin-T, and DINOv2. Interestingly, the federated transformer variations outperform the CNN ensemble (ResNet50 + VGG16) with an accuracy of 98.15%, FedDistill-DeiT achieving 97.79%, FedAvg-Swin-T reaching 97.75%, and FedDistill-Swin-T achieving a high ROC-AUC of 0.9977. These findings show that, even in the presence of diverse data distributions, federated Vision Transformers provide a reliable, scalable, and interpretable malaria screening solution that combines high accuracy with solid privacy guarantees.