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

Capture Global Feature Statistics for One-Shot Federated Learning

AAAI 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated learning with pretrained models · 9 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 FedCGS, which aggregates class-wise feature sums, counts and a global second-moment matrix from frozen pre-trained backbones into a parameter-free Gaussian Naive Bayes head for training-free one-shot FL, plus a personalized variant that uses the global prototypes as a regularizer.

Problem

One-shot federated learning suffers from high server/client computation cost and unstable performance under non-IID data; existing methods (distillation, generative, Bayesian) are costly or heterogeneity-sensitive, and GMM-sampling-based pre-trained approaches like FedPFT are limited by sampling var

Approach

Clients upload per-class feature sums, class counts and the sum of feature outer products computed with a frozen pre-trained extractor; the server reconstructs exact global mean and covariance (heterogeneity-invariant since they are sums), and builds a closed-form Gaussian Naive Bayes classifier head with no training. For personalization, clients download the fixed global prototypes once and use t

Key results

• On label-shift one-shot FL with alpha=0.05, FedCGS reaches 63.95% on CIFAR10 and 57.77% on SVHN, beating the best baseline by 7.87 and 15.22 points respectively (per the paper's text).
• Global statistics recovered by FedCGS deviate from centralized ground truth by only about 1e-7 to 1e-5 in L2 norm, insensitive to client count and heterogeneity.
• In feature-shift (domain generalization) on PACS and OfficeHome, FedCGS averages 62.14% and 60.12% versus FedPFT's 57.36% and 57.68%.

Measured effects reported in the paper

Surprising findings

• FedAvg-FT, a simple fine-tuning baseline, achieves the best personalized results on all datasets, outperforming many personalized FL methods including FedCGS.
• In the personalized setting, FedCGS with fixed global prototypes performs similarly to FedProto, which updates prototypes every round.

Limitations

• Designed primarily for image classification; extension to generation tasks is unexplored.
• Performance depends on the quality of the pre-trained model's features.
• Uploaded statistics still include per-class counts that need secure aggregation to protect label information.

Experimental setup

Datasets: CIFAR10 CIFAR100 SVHN PACS OfficeHome

Benchmarks: CIFAR10 CIFAR100 SVHN PACS OfficeHome

Metrics: test accuracy L2 deviation of global statistics PSNR for inversion attack

Baselines: FedAvg (one-shot) Ensemble DENSE Co-Boosting FedPFT Local-only FedAvg-FT FedProto

Builds on: FedPFT CCVR FedProto DENSE Co-Boosting

Closest papers by meaning

• FedLPA: One-shot Federated Learning with Layer-Wise Posterior Aggregation (NeurIPS 2024 · similarity 0.88)
• FedBEns: One-Shot Federated Learning based on Bayesian Ensemble (ICML 2025 · similarity 0.88)
• One-Shot Heterogeneous Federated Learning with Local Model-Guided Diffusion Models (ICML 2025 · similarity 0.88)
• One-shot Federated Learning via Synthetic Distiller-Distillate Communication (NeurIPS 2024 · similarity 0.88)
• FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.87)
• Does One-shot Give the Best Shot? Mitigating Model Inconsistency in One-shot Federated Learning (ICML 2025 · similarity 0.86)