Researchers propose FedKT-CSD, a framework for one-shot federated learning that achieves rigorous privacy guarantees while maintaining model quality across heterogeneous client data. The method leverages publicly pretrained autoencoders to create a shared latent space, allowing clients to encode private data and transmit only class-conditional latent statistics.

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

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