The Gaussian-Head OFL Family: One-Shot Federated Learning from Client Global Statistics

Researchers introduce Gaussian-Head OFL, a one-shot federated learning method that reduces communication overhead to a single round by transmitting only statistical summaries instead of full models. The approach combines closed-form Gaussian classifiers with synthetic data generation, achieving competitive accuracy while maintaining privacy and eliminating dependency on public datasets.

The Gaussian-Head OFL framework addresses a fundamental inefficiency in federated learning: the communication burden and privacy exposure from iterative model exchanges. Traditional federated learning requires multiple rounds of client-server communication, creating bottlenecks and increasing attack surface for sensitive model information. This research demonstrates that sufficient statistics—per-class counts and statistical moments—contain enough information to reconstruct effective classifiers without transmitting entire models, fundamentally reducing communication complexity from O(n) rounds to O(1).

The breakthrough emerges from assuming class-conditional Gaussianity in pretrained embeddings, a reasonable assumption for modern deep learning models. Rather than treating this as a limitation, the authors leverage it constructively through three complementary components: NB/LDA/QDA heads computed directly from statistics, FisherMix for linear classifiers trained on synthetic samples, and Proto-Hyper for refinement through knowledge distillation. This layered approach handles heterogeneous client models and non-IID data distributions without requiring public datasets—common constraints that limit practical deployment of existing one-shot methods.

For the distributed machine learning and edge computing sectors, this work meaningfully advances feasibility of federated systems in resource-constrained and privacy-sensitive environments. Medical institutions, financial firms, and IoT networks could deploy federated learning with significantly reduced communication costs and privacy risks. The data-free aspect eliminates regulatory and logistical complications around public data availability. The demonstrated robustness under non-IID conditions suggests real-world applicability where client data distributions diverge naturally.

Future validation should focus on performance across diverse model architectures and dataset characteristics beyond tested benchmarks, and quantitative privacy guarantees beyond the inference that sufficient statistics are less informative than full models.

• →One-shot federated learning reduces server-client communication from multiple rounds to a single exchange, dramatically lowering overhead
• →Transmitting only statistical summaries instead of full models enhances privacy while eliminating model homogeneity requirements
• →The method achieves state-of-the-art performance on non-IID data without requiring public datasets or additional information uploads
• →Combines classical statistical classifiers with modern synthetic data generation for improved robustness and accuracy
• →Framework addresses practical deployment barriers in federated learning for healthcare, finance, and edge computing applications