Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedSA: A Unified Representation Learning via Semantic Anchors for Prototype-based 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 FedSA, a prototype-based federated learning framework that uses embedding-projected class anchors as shared semantic prototypes with anchor-based regularization, margin-enhanced contrastive learning, and classifier calibration to counter statistical and model heterogeneity.
Problem
Prototype-based federated learning collects prototypes from biased local models with heterogeneous data and architectures, causing representation inconsistency, classifier divergence, and skewed prototype alignment that degrade client performance.
Approach
Decouple prototype generation from local representation learning by sharing server-side semantic anchors (pre-defined class anchors projected through a learned embedding layer) that guide local feature extractors and calibrate classifiers, and update them via EMA from global prototypes.
Key results
- Under statistical heterogeneity in the cross-device setting, FedSA reaches 88.78% on CIFAR-10, 48.42% on CIFAR-100 and 30.82% on Tiny-ImageNet, the best among compared methods.
- On CIFAR-100 with model heterogeneity (HtFE8), FedSA reaches 35.68% versus 16.31% for FedProto and 32.04% for FedTGP.
- Ablation on CIFAR-100 HtFE8 shows the full FedSA components raise accuracy from 29.54% to 35.68%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use semantic anchors with FedSA instead of the strongest baseline | FedSA versus strongest baseline · Cifar10, Cifar100, and Tiny-ImageNet, cross-silo and cross-device, statistical heterogeneity | test accuracy (%): 90.76$\pm$0.05 -> 90.88$\pm$0.04; 53.71$\pm$0.13 -> 54.39$\pm$0.12; 37.14$\pm$0.16 -> 37.30$\pm$0.13; 87.98$\pm$0.16 -> | Helped |
| Use FedSA rather than the strongest baseline across heterogeneous feature extractors | FedSA versus strongest baseline · Cifar100, HtFE_2, HtFE_3, HtFE_5, and HtFE_8 | test accuracy (%): 42.19$\pm$0.09 -> 43.54$\pm$0.21; 40.92$\pm$0.36 -> 41.75$\pm$0.16; 39.98$\pm$0.13 -> 40.96$\pm$0.12; 35.26$\pm$0.48 -> | Helped |
| Add embedding projection (ablation) | FedSA with EMA and anchor-based regularization · Cifar100, HtFE_8, cross-device | test accuracy (%): 29.54 -> 30.80 (+1.26) | Helped |
| Add classifier calibration to the embedding-projection variant (ablation) | FedSA ablation · Cifar100, HtFE_8, cross-device | test accuracy (%): 30.80 -> 30.61 | Hurt |
| Add margin-enhanced contrastive learning with embedding projection (ablation) | FedSA ablation · Cifar100, HtFE_8, cross-device | test accuracy (%): 29.54 -> 32.82 (+3.28) | Helped |
| Add classifier calibration to the embedding-and-MCL variant (ablation) | FedSA · Cifar100, HtFE_8, cross-device | test accuracy (%): 32.82 -> 35.68 | Helped |
| Set lambda_2 to 0.1 rather than 1 (ablation) | FedSA · Cifar100, HtFE_8, cross-device | test accuracy (%): 35.68 -> 33.42 | Hurt |
| Set lambda_2 to 5 rather than 1 (ablation) | FedSA · Cifar100, HtFE_8, cross-device | test accuracy (%): 35.68 -> 35.54 | Hurt |
| Set lambda_2 to 10 rather than 1 (ablation) | FedSA · Cifar100, HtFE_8, cross-device | test accuracy (%): 35.68 -> 35.12 | Hurt |
| Set lambda_3 to 0.1 rather than 1 (ablation) | FedSA · Cifar100, HtFE_8, cross-device | test accuracy (%): 35.68 -> 33.26 | Hurt |
| Set lambda_3 to 5 rather than 1 (ablation) | FedSA · Cifar100, HtFE_8, cross-device | test accuracy (%): 35.68 -> 34.66 | Hurt |
| Set lambda_3 to 10 rather than 1 (ablation) | FedSA · Cifar100, HtFE_8, cross-device | test accuracy (%): 35.68 -> 34.29 | Hurt |
Surprising findings
- FedProto collapses under heavier model heterogeneity, dropping to 16.31% on CIFAR-100 HtFE8, while knowledge-based baselines such as FedKD remain stable.
Limitations
- Hyperparameters lambda_2 and lambda_3 are sensitive; large values degrade accuracy.
- Evaluated only on image classification with small datasets (CIFAR-10/100, Tiny-ImageNet).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet
Metrics: test accuracy
Baselines: LG-FedAvg FedGen FML FedKD FedDistill FedProto FedTGP
Builds on: FedProto FedTGP FedRep FedRoD
Closest papers by meaning
- Beyond Federated Prototype Learning: Learnable Semantic Anchors with Hyperspherical Contrast for Domain-Skewed Data (AAAI 2025 · similarity 0.92)
- Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains (NeurIPS 2024 · similarity 0.90)
- Federated Model Heterogeneous Matryoshka Representation Learning (NeurIPS 2024 · similarity 0.88)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.87)
- Overcoming Data and Model heterogeneities in Decentralized Federated Learning via Synthetic Anchors (ICML 2024 · similarity 0.87)
- Beyond Prototypes: Semantic Anchor Regularization for Better Representation Learning (AAAI 2024 · similarity 0.87)
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.