Collaborative Synthetic Data Generation for Knowledge Transfer in Federated Learning

One-shot federated learning (OSFL) aims to reduce communication overhead by limiting client-server interaction to a single round. However, existing OSFL methods often struggle to simultaneously achieve high model utility, robustness to heterogeneous data distributions, and formal privacy guarantees. This paper asks: can we design a communication-efficient, privacy-preserving OSFL framework that avoids on-device training and complex parameter aggregation?

The authors propose FedKT-CSD, a framework that leverages a frozen, publicly pretrained autoencoder to map private client data into a shared latent space. Instead of sharing model parameters or raw data, each client performs a single forward pass to encode its images and computes simple class-conditional statistics (sums and outer products of latent vectors). These lightweight summaries are transmitted to the server via secure aggregation, where calibrated Gaussian noise is added to satisfy $(\varepsilon,\delta)$-differential privacy. The server then uses these noisy statistics to reconstruct a synthetic dataset by sampling from the recovered latent distributions and decoding them back into pixel space.

FedKT-CSD provides formal privacy guarantees by design, as the autoencoder is frozen and the transmitted statistics have bounded sensitivity. The method is highly efficient, requiring only a single communication round and minimal client-side computation (no gradient descent or training). Despite the privacy constraints, the framework remains competitive with or outperforms non-private baselines across diverse datasets, including natural, satellite, and medical imagery. Furthermore, the synthetic data generated by the server is versatile and can be used for downstream tasks like personalized federated learning or model pretraining.

This work bridges the gap between communication-efficient federated learning and rigorous privacy. By shifting the focus from parameter sharing to latent-space statistics, FedKT-CSD enables deployment in resource-constrained environments where traditional multi-round federated learning is impractical. Its reliance on public, frozen models makes it a scalable solution for collaborative learning without the need for expensive on-device optimization.