One-Shot Federated Distillation Using Monoclass Teachers: A Study of Knowledge Fragmentation and Out-of-Distribution Supervision

Cedric Maron, Virginie Fresse, ORZALESI

Abstract: The performance of machine learning models critically depends on the quality and diversity of training data. However, privacy, legal, and proprietary concerns often limit direct data sharing. Many organizations possess high-quality data for specific classes and may wish to share the knowledge derived from it without revealing the data or engaging in collaborative training. While federated learning (FL) enables distributed model training, it typically assumes mutual benefit, requires repeated communication, and produces a shared global model. Another paradigm, knowledge distillation (KD), allows a student model to learn from teacher predictions. We propose a one-shot federated distillation method in which a single client learns from monoclass teacher models trained independently by multiple providers. Each provider shares its model once, and the client combines these with unlabeled data to distill a multiclass student model—aggregating knowledge from disjoint, class-specific sources. This unidirectional, asymmetric setup poses a key challenge: out-of-distribution (OOD) supervision, where monoclass teachers often mispredict unseen inputs, leading to noisy signals for the student. The main contribution of this work is a systematic study of knowledge fragmentation in one-shot federated distillation with monoclass teachers. We evaluate five configurations with varying class coverage per provider and show that increasing fragmentation intensifies OOD supervision, degrading student performance. Experiments on MNIST, FashionMNIST, and CIFAR-10 confirm that fragmentation consistently reduces student accuracy. To mitigate this, we discuss three strategies: (1) exposing teachers to diverse off-class examples, (2) penalizing overconfidence, and (3) using contrastive learning to sharpen feature boundaries.

Submission Length: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Gang_Niu1

Submission Number: 4845