An Interpretable Federated Learning Framework for Secure Genomics

Interpretability and data security are critical requirements in federated learning, particularly in healthcare and genomics where sensitive patient data is involved. This research presents a privacy preserving and interpretable federated learning framework using the Nash equilibrium from noncooperative game theory. Agents (e.g., hospitals or research centers) decide how much local genomic data to share by optimizing utility functions that balance model performance with privacy protection. To improve interpretability, the framework incorporates graphbased methods: sparse inverse covariance estimation (SICE) and Markov random fields (MRFs). SICE identifies conditional gene dependencies by estimating a sparse inverse covariance matrix. MRFs visualize direct gene-gene interactions through network structures. These techniques help explain how key features are selected and how agent level decisions impact the global model. Experiments on gene expression datasets for Alzheimer's disease, breast cancer, and pancreatic cancer confirm the model's ability to identify consistent and biologically meaningful gene sets across agents. The results demonstrate strong model performance, robust privacy guarantees, and transparent decision making.