Research map / Federated learning and privacy / Personalized and heterogeneous FL
FOOGD: Federated Collaboration for Both Out-of-distribution Generalization and Detection
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 FOOGD, a federated framework that jointly handles OOD detection and OOD generalization by estimating client score models via score matching with MMD (SM-MMD) and regularizing feature extractors with a kernelized Stein augmentation (SAG).
Problem
Federated learning deployed on wild data must simultaneously handle covariate-shift OOD (generalization) and semantic-shift OOD (detection) under non-IID client distributions, but prior FL work addresses only one of these.
Approach
Each client trains a feature-space score model with a mix of denoising score matching and MMD matching on Langevin-generated samples; the aggregated score model gives a global density guide, whose score norm detects OUT data, while a kernelized Stein discrepancy between original and Fourier-augmented features regularizes the feature extractor for IN-C generalization.
Key results
- On CIFAR-10 (alpha=0.1), FedAvg+FOOGD raises AUROC on LSUN-C from 58.05 (FedAvg) to 91.21 and ACC-IN-C from 65.44 to 73.71.
- FedRoD+FOOGD reaches 92.74 ACC-IN-C and 91.76 AUROC on CIFAR-10 (alpha=0.1), the best among compared methods in the table.
- On PACS leave-one-domain-out, FedRoD+FOOGD averages 83.53 versus 77.40 for FedTHE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FOOGD to FedAvg | FedAvg · CIFAR-10 α=0.1 | ACC-IN-C: 65.44 -> 73.71 | Helped |
| Add FOOGD to FedAvg | FedAvg · CIFAR-10 α=0.1 | FPR95: 83.41 -> 35.32 | Helped |
| Add FOOGD to FedRoD | FedRoD · CIFAR-10 α=0.1 | AUROC: 80.96 -> 91.76 | Helped |
| Add FOOGD to FedRoD | FedRoD · CIFAR-100 α=0.1 | ACC-IN-C: 69.26 -> 75.70 | Helped |
| Add FOOGD to FedAvg | FedAvg · PACS Average | accuracy: 71.52 -> 79.92 | Helped |
| Add FOOGD to FedRoD | FedRoD · PACS Average | accuracy: 75.39 -> 83.53 | Helped |
| Add SAG (ablation) | FedAvg+FOOGD · CIFAR-10 α=0.1 | ACC-IN-C: 70.79 -> 73.71 | Helped |
| Add SAG (ablation) | FedAvg+FOOGD · CIFAR-100 α=0.1 | ACC-IN-C: 48.35 -> 51.69 | Helped |
| Add technameA (score matching with MMD) (ablation) | FedAvg+FOOGD · CIFAR-10 α=0.1 | FPR95: 41.86 -> 35.32 | Helped |
| Add technameA (score matching with MMD) (ablation) | FedAvg+FOOGD · CIFAR-100 α=0.1 | AUROC: 89.26 -> 91.41 | Helped |
| Train the feature backbone instead of fixing it (ablation) | FedAvg+FOOGD · CIFAR-10 α=0.1 | ACC-IN-C: 65.44 -> 73.71 | Helped |
Surprising findings
- Personalized FL methods generally detect OUT data worse than non-personalized ones, suggesting a conflict between detection and personalized prediction under non-IID data.
- Score-based density estimation alone (lambda_m=0) and MMD alone (lambda_m=1) are both less compact than a mix at lambda_m=0.1 in the 2-D toy example.
Limitations
- Privacy exposure is only argued qualitatively; differential privacy is left as future work.
- Evaluation limited to image benchmarks (CIFAR, TinyImageNet, PACS) with WideResNet and ResNet18 backbones.
- Theoretical error bound depends on an assumed MMD upper bound C and a loose Jensen-based aggregation argument.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet CIFAR-10-C CIFAR-100-C TinyImageNet-C PACS
Benchmarks: CIFAR-10 CIFAR-100 TinyImageNet SVHN Texture iSUN LSUN-C LSUN-R PACS
Metrics: ACC-IN ACC-IN-C AUROC FPR95
Baselines: FedAvg FedRoD FOSTER FedLN FedATOL FedT3A FedIIR FedTHE FedICON
Builds on: FedAvg FedRoD SCONE FOSTER Fourier augmentation
Closest papers by meaning
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.87)
- FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.87)
- Benchmarking Algorithms for Federated Domain Generalization (ICLR 2024 · similarity 0.87)
- FedImpro: Measuring and Improving Client Update in Federated Learning (ICLR 2024 · similarity 0.86)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.86)
- FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.