SFL-MTKD++: Split ViT Federated Learning with Adaptive Multi-Teacher Distillation for Interpretable Malaria Diagnosis

Stringent data privacy regulations and domain heterogeneity inhibit centralized deep learning for malaria diagnosis. We introduce SFL-MTKD++, a Split Federated Learning framework augmented with Adaptive Multi-Teacher Knowledge Distillation. By leveraging a partitioned Vision Transformer with a secure feature bottleneck, the architecture facilitates collaborative training across isolated silos while strictly precluding raw data exposure. On a harmonized multi-source dataset, our method achieves 97.64\% accuracy and a 0.9948 ROC-AUC, significantly outperforming existing federated baselines. To ensure clinical reliability, we integrate LIME and saliency map visualizations, confirming that the model identifies pathologically relevant features rather than artifacts. This work establishes a robust paradigm for deploying high-fidelity, interpretable AI in resource-constrained healthcare settings without compromising patient confidentiality.