Federated Synthetic Data Generation

A generator network—typically a GAN or VAE—is trained across multiple silos without pooling raw data. Each client trains a local discriminator or encoder on its private dataset, and only model gradients or parameters are transmitted to a central aggregation server. The global generator learns to model the joint data distribution of all participants without ever seeing a single real record. This architecture preserves the statistical correlations, rare edge cases, and multimodal distributions present across the combined datasets while maintaining strict data locality.

Synthetic data alone does not guarantee privacy—models can memorize and regurgitate training samples. Federated synthetic generators integrate Differential Privacy (DP) by clipping gradient norms and injecting calibrated Gaussian noise during local training. This provides a mathematically provable bound (ε, δ) on the privacy leakage of any single individual's record. The resulting synthetic dataset carries a privacy guarantee that the presence or absence of any patient in the original silos cannot be reliably inferred, enabling downstream sharing with researchers who lack direct data access.

The primary challenge is maintaining high utility in the generated data. Federated architectures employ techniques such as federated GAN discriminators that learn to distinguish real from synthetic samples locally, then share their learned feature representations. The aggregated global discriminator guides the generator to produce samples that are indistinguishable from real data across all sites. Evaluation metrics include marginal distribution similarity, joint correlation preservation, and downstream model utility—ensuring a classifier trained on synthetic data performs comparably to one trained on the original combined dataset.

Hospital datasets are inherently non-IID—different patient demographics, equipment vendors, and clinical practices create statistical heterogeneity. Federated synthetic generation must account for this by learning site-specific latent representations or conditioning the generator on site metadata. Advanced approaches use conditional GANs where the generator receives a site identifier, allowing it to produce synthetic cohorts that faithfully represent each institution's unique distribution while still benefiting from the broader statistical patterns learned across the federation.

Even model updates can leak information through gradient inversion attacks. Federated synthetic data pipelines employ secure aggregation protocols that encrypt individual client updates using pairwise masking or threshold secret sharing. The central server can only compute the sum of all updates—never inspecting any single institution's contribution. Combined with DP noise, this provides defense-in-depth: even if the aggregation server is compromised, individual training samples cannot be reconstructed from the aggregated gradient information.

Evaluating synthetic data quality typically requires comparison against real holdout sets—impossible when data cannot be centralized. Federated validation uses distributed metrics computation: each site computes local statistical measures (e.g., pairwise correlation matrices, univariate histograms) on its real data and the synthetic samples, then shares only these aggregate statistics with the coordinator. Techniques like federated maximum mean discrepancy allow quantitative assessment of distributional similarity without exposing patient-level information.

In real-world federated settings, data across silos is non-identically and independently distributed. One hospital may have predominantly geriatric patients while another is a pediatric center. This statistical heterogeneity can cause a centrally aggregated generative model to diverge or produce low-quality synthetic data. Federated synthetic data generation algorithms must employ specialized techniques to handle this skew.

• FedProx: Adds a proximal term to local training to constrain updates.
• Clustered FL: Groups clients with similar distributions before training separate generators.