Research map / Federated learning and privacy / Communication-efficient FL
Covariances for Free: Exploiting Mean Distributions for Training-free Federated Learning
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 FedCOF, a training-free federated classifier initialization that estimates class covariances unbiasedly from client class means only, and initializes a classifier using within-class covariances.
Problem
Training-free federated learning with pre-trained models: second-order statistics (Fed3R) improve classifier initialization but incur large communication and privacy costs, while class-mean-only methods (FedNCM) are weaker.
Approach
Derives a provably unbiased estimator of class covariance matrices from per-client class means (between-client variance of means scaled by client sample counts), adds shrinkage, and uses only within-class covariances to build a ridge-style classifier initialization, so only means and counts are communicated.
Key results
- On iNat-120K with ViT-B/16, FedCOF reaches 63.1% vs 62.5% for Fed3R at 167.7 MB versus Fed3R's 22050.2 MB communication.
- On Cars with ViT-B/16, FedCOF improves over FedNCM by 26% (52.5 vs 26.2) at identical communication cost.
- Compared with federated prompt-tuning (PFPT) using ViT-B/32 on CUB200, FedCOF achieves 65.0% vs 38.6% with 7.2 MB vs 1765.5 MB communication.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedCOF instead of Fed3R | training-free federated classifier initialization · CUB200, Cars, and iNat-120K | accuracy: 3.3%, 3.9% and 2.2% on CUB200; 4.1% to 5.7% on Cars; 0.4%, 2.6% and 0.6% on iNat-120K | Helped |
| Use FedCOF instead of Fed3R | training-free federated classifier initialization · CIFAR-100 and ImageNet-R | accuracy | Mixed |
| Use FedCOF instead of FedNCM with the same communication budget | training-free federated classifier initialization · Cars | accuracy: 24% using SqueezeNet and 26% using ViT-B/16 | Helped |
| Use FedCOF instead of federated prompt-tuning methods | ViT-B/32 training-free classification · CIFAR-100, ImageNet-R, CUB200, and Cars | accuracy | Helped |
| Remove between-class scatter from classifier initialization (ablation) | FedCOF · CIFAR-100 | accuracy: 57.1 -> 57.3 | Helped |
| Remove between-class scatter from classifier initialization (ablation) | FedCOF · ImageNet-R | accuracy: 37.6 -> 38.6 | Helped |
| Remove between-class scatter from classifier initialization (ablation) | FedCOF · CUB200 | accuracy: 50.4 -> 53.7 | Helped |
| Remove between-class scatter from classifier initialization (ablation) | FedCOF · Cars | accuracy: 41.4 -> 44.8 | Helped |
| Remove between-class scatter while retaining within-class scatter (ablation) | federated classifier initialization · CIFAR-100 | accuracy: 56.3 -> 56.3 | No clear effect |
| Remove between-class scatter while retaining within-class scatter (ablation) | federated classifier initialization · ImageNet-R | accuracy: 36.8 -> 37.2 | Helped |
| Remove between-class scatter while retaining within-class scatter (ablation) | federated classifier initialization · CUB200 | accuracy: 51.6 -> 53.5 | Helped |
| Remove between-class scatter while retaining within-class scatter (ablation) | federated classifier initialization · Cars | accuracy: 42.4 -> 44.6 | Helped |
Surprising findings
- Removing between-class scatter from the ridge regression classifier improves accuracy in both centralized and federated settings (e.g., CUB200 federated 49.7 with between-class only vs 53.5 with within-class only using SqueezeNet).
- Under feature-shift non-iid settings on DomainNet, Fed3R (81.9%) outperforms FedCOF (74.1%) with the same class-mean-based estimator's bias noted.
Limitations
- Estimator quality depends on the number of clients; performance drops with few clients unless multiple means per client are sampled.
- Unbiasedness relies on iid features per class across clients; bias appears under feature shift and Fed3R outperforms FedCOF there.
Experimental setup
Datasets: CIFAR-100 ImageNet-R CUB200 Stanford Cars iNaturalist iNaturalist-Users-120K DomainNet
Benchmarks: CIFAR-100 ImageNet-R CUB200 Stanford Cars iNat-120K DomainNet
Metrics: accuracy communication cost (MB)
Baselines: FedNCM Fed3R FedCOF Oracle CCVR FedAvg-PT FedProx-PT PFPT FedAvg FedAdam FedAvg-LP
Builds on: FedNCM Fed3R CCVR Linear Discriminant Analysis
Closest papers by meaning
- Federated Learning with Sample-level Client Drift Mitigation (AAAI 2025 · similarity 0.87)
- FedUV: Uniformity and Variance for Heterogeneous Federated Learning (CVPR 2024 · similarity 0.86)
- Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · similarity 0.86)
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.86)
- TurboSVM-FL: Boosting Federated Learning through SVM Aggregation for Lazy Clients (AAAI 2024 · similarity 0.86)
- FedAdaVR: Adaptive Variance Reduction for Robust Federated Learning under Limited Client Participation (NeurIPS 2026 · 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.