Wenbo Yu, Qingqing Ye, Sen Zhang, Junxu Liu, Haibo Hu · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3836289
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Deploying Diffusion Models (DMs) in federated settings has recently attracted growing attention for their ability to harness the strong generative capacity of diffusion models. However, their practical deployment remains hindered by privacy risks, as sharing local models can potentially leak sensitive training data. To mitigate this, we propose PrivFDM, a differentially private federated diffusion model for image synthesis that generates high-fidelity synthetic images while ensuring rigorous privacy guarantees. PrivFDM presents a local pre-training method with autoencoder?based data perturbation that injects DP noise directly into the low?dimensional representations of the training data. This design reduces data sensitivity and significantly lowers privacy budget consumption. To support effective and private learning under heterogeneous client data, we further develop a local fine-tuning strategy with model perturbation and common knowledge sharing. By integrating DP?SGD and conditioning diffusion models through label embeddings, our approach mitigates the disruptive effect of client divergence and enhances the utility of the synthesized images. We theoretically prove that PrivFDM ensures DP guarantees across the entire federated training pipeline. Extensive experiments on multiple datasets demonstrate the strong performance and practical advantages of the proposed PrivFDM in federated learning.
No comments yet — start the discussion below.