Interpretable and Privacy-Preserving Federated Learning via Subspace Representations

Federated Learning (FL) has emerged as a vital paradigm for privacy-preserving collaborative learning. While prototype-based interpretability offers explanations in centralized settings, its extension to FL is hindered by the risk of leaking sensitive prototypical patches and the challenge of aggregating heterogeneous client concepts. We propose FedGraSP, a novel manifold-aware, subspace-based framework that aggregates class evidence on the Grassmann manifold for part-level interpretable FL. By encoding representative part-level prototypes as subspaces on the Grassmann manifold and exploiting its rotation invariance, FedGraSP achieves well-generalized collaborative learning without transmitting private raw image patches. Our framework utilizes a projection-retraction-based gradient update to maintain manifold obedience and extends naturally to personalized FL by repurposing the final layer to adapt to each user's dataset distribution. Extensive evaluations on fine-grained classification datasets demonstrate that FedGraSP provides faithful, part-level explanations while maintaining high utility, bridging the gap between transparent reasoning and data minimization in decentralized environments.