Research map / Federated learning and privacy / Personalized and heterogeneous FL

Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models

AAAI 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated semi-supervised and graph learning · 28 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FedDISC, a one-shot semi-supervised federated learning method that uploads clustered CLIP feature centroids and domain representations from clients, and uses a pre-trained Stable Diffusion model on the server to generate labeled synthetic data for training a CLIP-based classifier.

Problem

Semi-supervised federated learning (labeled server data, unlabeled client data) suffers from multi-round communication costs, non-IID data heterogeneity, and heavy on-device training burdens.

Approach

Clients do no training; they send noise-perturbed cluster centroids and per-category domain-specific feature averages, and the server conditions a pre-trained diffusion model on them to generate pseudo-labeled synthetic data that mimics client distributions for training a global classifier in a single communication round.

Key results

• FedDISC achieves average accuracy of 59.42 on OpenImage and 53.68 on DomainNet, the best among non-ceiling methods, in one communication round.
• On NICO++_U, FedDISC averages 91.03 versus 78.09 for SemiFL and 88.67 for Prompt, and it surpasses the centralized ceiling on some clients (e.g., NICO++_C client0 87.97 vs 89.19 ceiling is lower, but DomainNet clipart 72.54 vs 81.54 is lower too; surpassing ceiling occurs on NICO++_U-like client com
• Uploads about 4.23M parameters versus 30x632.08M for FedAvg, and client compute of 334.73 Gflops versus 30x1004.19 Gflops for FedAvg.

Measured effects reported in the paper

Surprising findings

• Generated synthetic data sometimes matches or exceeds centralized ceiling performance on certain client domains, despite using no client training.
• Zero-shot CLIP outperforms Fine-tune on some clients, showing fine-tuning on pseudo-labeled centroids can be weaker than zero-shot.

Limitations

• Stable Diffusion has limited ability to render text in images, leading to poor performance on DomainNet infograph.
• Privacy protection is argued qualitatively via visualizations rather than formally quantified.

Experimental setup

Datasets: DomainNet OpenImage NICO++

Benchmarks: DomainNet OpenImage NICO++

Metrics: classification accuracy

Baselines: Fine-tune Zero-shot FedAvg SemiFL RSCFed Prompt Ceiling (centralized training)

Builds on: FedAvg SemiFL RSCFed Stable Diffusion CLIP

Closest papers by meaning

• One-Shot Heterogeneous Federated Learning with Local Model-Guided Diffusion Models (ICML 2025 · similarity 0.91)
• (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.91)
• SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.91)
• Progressive Distribution Matching for Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.91)
• BSemiFL: Semi-supervised Federated Learning via a Bayesian Approach (ICML 2025 · similarity 0.89)
• Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.88)