FedKT-CSD enables one-shot federated learning with formal differential privacy

The authors propose FedKT-CSD, a framework for one-shot federated learning that addresses the challenge of maintaining model quality and privacy when client data distributions diverge. The method uses publicly pretrained autoencoders to create a shared latent space where clients encode private data and transmit class-conditional statistics.

• Clients perform a single forward pass to compute latent statistics, keeping client-side computation lightweight.
• The server aggregates these statistics via secure aggregation and adds calibrated differential privacy noise.
• A synthetic dataset is decoded from the aggregated statistics for training a global model.
• The design provides formal $(\varepsilon,\delta)$-differential privacy by construction.

FedKT-CSD scales to a large number of clients and is competitive with or outperforms non-private baselines across diverse datasets and heterogeneity settings.