Andrea Moleri, Christian Internó, Ali Raza, Markus Olhofer, Barbara Hammer, "Geodesic-SVDD Latent Mixup for Distributed Learning", International Joint Conference on Neural Networks (IJCNN 2026), 2026.

Federated Learning (FL) offers a privacy-enhancing paradigm for distributed learning; however, training One-Class Classifiers (OCC) such as Deep Support Vector Data Description (Deep SVDD) in distributed settings remains challenging due to statistical heterogeneity (Non-IID data) and the risk of local mode collapse. This paper presents a novel Federated Deep SVDD framework designed to enforce a compact, hyperspherical description of normal data across disjoint clients. We introduce a Residual Spherical Encoder, a specialized architecture utilizing residual blocks and strict L2 normalization to map input data onto a unit hypersphere. To mitigate the sparsity of local data and improve the continuity of the learned manifold, we propose Geodesic Mixup, a regularization technique that synthesizes latent samples using Spherical Linear Interpolation (SLERP) rather than Euclidean interpolation.