KDD 2026 - Communication-Efficient Federated Graph Classification via Generative Diffusion Modeling
Why It Matters
The approach makes large‑scale federated graph learning economically viable, unlocking privacy‑sensitive applications while slashing network costs.
Key Takeaways
- •Federated graph classification faces high communication overheads during training.
- •Proposed CFTC framework reduces communication to three rounds.
- •Uses generative diffusion models to synthesize static graphs locally.
- •Clients train local GNNs on real and generated graphs.
- •Aggregated model improves accuracy while preserving data privacy.
Summary
The paper presented at KDD 2026 tackles the problem of graph classification in a federated setting, where multiple clients must collaboratively train a model without sharing raw graph data.
Current federated graph neural networks (GNNs) suffer from two major issues: frequent, bandwidth‑heavy model exchanges and performance drift caused by heterogeneous graph structures across clients. To overcome these, the authors introduce CFTC, a four‑step framework that leverages generative diffusion models to capture the underlying distribution of each client’s graphs.
Each client first trains a diffusion‑based generator on its local graphs and sends the generator to the server. The server redistributes the generators, after which clients synthesize a set of static graphs, combine them with their real data, and train a local GNN. Only the trained GNN weights are uploaded, and the server aggregates them into a global model. The authors claim the entire process requires just three communication rounds and a single aggregation step.
By cutting communication to three rounds and using synthetic graphs to homogenize training data, CFTC promises lower bandwidth costs, faster convergence, and higher classification accuracy—benefits that could accelerate privacy‑preserving analytics in drug discovery, fraud detection, and other graph‑rich industries.
Comments
Want to join the conversation?
Loading comments...