Research map / Federated learning and privacy / Personalized and heterogeneous FL
Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains
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 FedPLVM, a federated prototype learning method for heterogeneous domains that uses dual-level (local and global) prototype clustering and an alpha-sparsity prototype loss to mitigate cross-domain representation variance.
Problem
Federated prototype learning with heterogeneous data domains creates one averaged prototype per class per client, which fails to capture intra-class variance, leading to unequal learning across domains (hard domains such as SVHN underperform) and communication/privacy costs.
Approach
Replace averaged local prototypes with multiple FINCH-clustered local prototypes per class, cluster again on the server for global prototypes, and train with an alpha-sparsity contrastive prototype loss (cosine similarity raised to power alpha in (0,1)) plus a correction term that pulls intra-class similarity toward 1.
Key results
- On Digit-5, FedPLVM reaches 69.25% average test accuracy vs 66.88% for FPL (+2.37), with SVHN improving to 42.08%.
- On Office-10, FedPLVM reaches 60.65% average accuracy vs 58.55% for FPL.
- On DomainNet (10-class subset), FedPLVM reaches 42.55% average accuracy vs 41.31... reported FedProto 40.99 and FPL 40.31; ablation shows dual clustering with global clustering reduces prototypes per client from 100.92 to 21.20 with comparable accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedPLVM instead of FPL | Federated prototype learning · Digit-5 | average test accuracy: 66.88 -> 69.25 | Helped |
| Use FedPLVM instead of FPL | Federated prototype learning · Office-10 | average test accuracy: 58.55 -> 60.65 | Helped |
| Use FedPLVM instead of FedProto | Federated prototype learning · DomainNet | average test accuracy: 40.99 -> 42.55 | Helped |
| Use local clustering rather than local averaging, retaining global clustering (ablation) | FedPLVM · Digit-5 | average test accuracy: 66.40 -> 69.48 | Helped |
| Add global prototype clustering rather than averaging global prototypes (ablation) | FedPLVM with local clustering · Digit-5 | average test accuracy: 68.11 -> 69.25 | Helped |
| Add global prototype clustering instead of distributing all local clustered prototypes (ablation) | FedPLVM · Digit-5 | average test accuracy: 69.18 ± 0.77 -> 69.47 ± 0.71 | No clear effect |
| Add the contrastive term to the prototype loss (ablation) | FedPLVM · Digit-5 | average test accuracy: +1.46% | Helped |
| Add the corrective term to the prototype loss (ablation) | FedPLVM · Digit-5 | average test accuracy: +0.99% | Helped |
| Combine the contrastive and corrective terms in the α-sparsity prototype loss (ablation) | FedPLVM · Digit-5 | average test accuracy: +2.29% | Helped |
| Set α within (0,1) rather than α=1 (ablation) | FedPLVM α-sparsity prototype loss · Digit-5 | test accuracy | Helped |
| Replace FINCH with K-Means using adaptive K (ablation) | FedPLVM dual-level clustering · Digit-5 | average test accuracy: 69.25 -> 68.97 | Hurt |
| Add differential privacy to local clustered prototypes (ablation) | FedPLVM with no privacy protection · Digit-5 | average test accuracy: 69.25 -> 68.48 | Hurt |
Surprising findings
- Directly transferring all local prototypes without clustering performs comparably to dual-level clustering but requires about 4.76x the communication cost.
- Averaging local prototypes with global clustering yields only small gains over fully averaged prototypes, suggesting local clustering is the key component.
Limitations
- Does not address noisy labels, which the authors note could distort prototypes.
- Experiments use very small per-client training sets (100 or 300 samples) and a limited number of clients and domains.
Experimental setup
Datasets: Digit-5 Office-10 DomainNet
Benchmarks: Digit-5 Office-10 DomainNet
Metrics: test accuracy average accuracy per domain
Baselines: FedAvg FedProx FedProto FedPCL FedFA FPL
Builds on: FedPL FPL FedProto FINCH FedAvg
Closest papers by meaning
- FedSA: A Unified Representation Learning via Semantic Anchors for Prototype-based Federated Learning (AAAI 2025 · similarity 0.90)
- Beyond Federated Prototype Learning: Learnable Semantic Anchors with Hyperspherical Contrast for Domain-Skewed Data (AAAI 2025 · similarity 0.88)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.88)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.88)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.88)
- FedGMKD: An Efficient Prototype Federated Learning Framework through Knowledge Distillation and Discrepancy-Aware Aggregation (NeurIPS 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.