Abstract
Federated learning (FL) enables multiple clients to train models collectively while preserving data privacy. However, FL faces challenges in terms of communication cost and data heterogeneity. One-shot federated learning has emerged as a solution by reducing communication rounds, improving efficiency, and providing better security against eavesdropping attacks. Nevertheless, data heterogeneity remains a significant challenge, impacting performance. This work explores the effectiveness of diffusion models in one-shot FL, demonstrating their applicability in addressing data heterogeneity and improving FL performance. Additionally, we investigate the utility of our diffusion model approach, FedDiff, compared to other one-shot FL methods under differential privacy (DP). Furthermore, to improve generated sample quality under DP settings, we propose a pragmatic Fourier Magnitude Filtering (FMF) method, enhancing the effectiveness of generated data for global model training.
Introduction
Standard Federated Learning (FL) typically requires numerous communication rounds between clients and a central server, leading to high latency and increased vulnerability to eavesdropping. One-shot FL addresses these issues by requiring only a single round of communication. However, it struggles with data heterogeneity (Non-IID data), where local models become biased toward their specific data distributions. Previous attempts to mitigate this using Generative Adversarial Networks (GANs) or Variational Autoencoders (VAEs) often failed to produce high-quality, diverse samples, especially under strict Differential Privacy (DP) constraints.
The FedDiff Framework
The proposed FedDiff framework shifts the generative paradigm to Diffusion Models (DMs). Unlike GANs or VAEs, DMs are more resilient to the noise introduced by DP-SGD during training.
The FedDiff process involves:
- Local Training: Each client trains a class-conditioned Diffusion Model on its private data using DP-SGD to ensure privacy.
- Model Aggregation: Clients transmit their generative models to the server in a single round.
- Data Synthesis: The server generates a synthetic dataset from the aggregated models. The slow sampling speed of DMs is mitigated here as it occurs offline on the server.
- Fourier Magnitude Filtering (FMF): To handle "blurry" images produced under high DP noise, clients share a privacy-preserved average of their data's Fourier magnitudes. The server uses this "style signature" to filter the generated samples, retaining only those that match the original data's structural characteristics.
Experimental Evaluation
FedDiff was evaluated on FashionMNIST, PathMNIST, and CIFAR-10 datasets, with data heterogeneity simulated via a Dirichlet distribution.
Key Results
- Accuracy Boost: On CIFAR-10 with high heterogeneity, FedDiff achieved ~57% accuracy, significantly outperforming FedCVAE (~34%).
- Privacy Resilience: Under strict privacy budgets ($\epsilon=10$), FedDiff maintained a substantial lead, reaching ~73% accuracy on FashionMNIST compared to less than 45% for other one-shot methods.
- FMF Effectiveness: The Fourier filtering method provided an additional 2% to 4% accuracy gain in high-noise DP settings by curating higher-quality synthetic data.
Discussion and Future Work
FedDiff demonstrates that Diffusion Models are uniquely suited for one-shot FL because their primary drawback—sampling speed—is irrelevant in a server-side synthesis context. Furthermore, their stable training objective makes them robust to DP noise.
While FedDiff is a significant advancement, challenges remain in preserving structure under extremely strict privacy budgets on complex datasets. Future research will explore focusing DMs on task-relevant features and developing DP algorithms specifically optimized for the denoising process.
Summary updated