Undergraduate Thesis – A Shared Embedding Framework for Interpretable Heterogeneous Federated Learning in Multimodal Breast Cancer Detection without Raw Data Sh

Designed a federated learning framework (PyTorch, Flower) unifying 3 structurally different breast-

cancer datasets into a shared embedding space; equal-weight aggregation improved the smallest

dataset’s accuracy by +11.8% (70.8% → 79.2%) over local-only training while holding 98.3–98.8%

accuracy on the larger nodes, with SHAP-based interpretability and differential-privacy-protected

aggregation.