Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Generalized Category Discovery
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.
Introduces Fed-GCD, a federated setting for generalized category discovery, and proposes AGCL, which uses learnable local and global Gaussian mixture models (aggregated on the server by client semantics association) for contrastive learning, evaluated on six image datasets.
Problem
Generalized category discovery requires centralized access to data, which is infeasible when data are distributed across privacy-constrained clients with partially labeled data and heterogeneous, non-overlapping label spaces.
Approach
Model each class as a learnable Gaussian mixture component trained with a Mahalanobis-distance contrastive loss plus a maximum-likelihood regularizer; clients upload local GMMs, the server aggregates them via sampling and FINCH re-clustering into a global GMM, and clients jointly use local and global GMM contrastive losses.
Key results
- On CUB-200 under the EH setting (beta=0.05), FedAvg + AGCL reaches 53.1 All accuracy versus 43.3 for FedAvg + GCD baseline.
- On CIFAR-10 under NH setting, FedAvg + AGCL achieves 84.7 All accuracy versus 80.7 for FedAvg + GCD.
- Ablation on CUB-200 shows local GCL improves new-class accuracy by 10.9% over the GCD baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add local GMM contrastive learning to FedAvg + GCD (ablation) | Fed-GCD · CUB-200 and Oxford-Pet EH | All accuracy: 43.3 -> 52.2; 72.1 -> 79.5 | Helped |
| Add the GMM maximum-likelihood regularization (ablation) | Fed-GCD with local GCL · CUB-200 and Oxford-Pet EH | All accuracy: 48.9 -> 50.6; 76.8 -> 78.0 | Helped |
| Add global GMM contrastive learning through CSA (ablation) | Fed-GCD with local GCL · CUB-200 and Oxford-Pet EH | All accuracy: 52.2 -> 53.1; 79.5 -> 81.4 | Helped |
| Use FedProx instead of FedAvg with AGCL | Fed-GCD · CIFAR-10, CIFAR-100, ImageNet-100, CUB-200, Stanford Cars, and Oxford-Pet; NH and EH | All accuracy: 84.7 -> 84.8; 56.1 -> 55.9; 74.8 -> 74.7; 82.5 -> 83.0; 54.2 -> 54.7; 73.1 -> 74.2; 55.2 -> 55.4; 38.2 -> 38.5; 82.7 -> | Mixed |
| Use AGCL instead of FedAvg + GCL (ablation) | Fed-GCD · CIFAR-10, CIFAR-100, and ImageNet-100 NH | All accuracy: 83.2 -> 84.7; 54.1 -> 56.1; 74.1 -> 74.8 | Helped |
| Use AGCL instead of FedAvg + GCL (ablation) | Fed-GCD · CIFAR-10, CIFAR-100, and ImageNet-100 EH | All accuracy: 82.2 -> 82.5; 52.1 -> 54.2; 72.5 -> 73.1 | Helped |
| Use AGCL instead of FedAvg + GCL (ablation) | Fed-GCD · CUB-200, Stanford Cars, and Oxford-Pet NH | All accuracy: 53.7 -> 55.2; 36.0 -> 38.2; 80.7 -> 82.7 | Helped |
| Use AGCL instead of FedAvg + GCL (ablation) | Fed-GCD · CUB-200, Stanford Cars, and Oxford-Pet EH | All accuracy: 52.2 -> 53.1; 35.3 -> 36.4; 79.5 -> 81.4 | Helped |
| Use AGCL instead of FedAvg + GCD with 10 clients (ablation) | Fed-GCD · CIFAR-10, CIFAR-100, and ImageNet-100 | All accuracy: 63.4 -> 68.1; 47.3 -> 52.2; 62.3 -> 67.5 | Helped |
| Increase the number of clients to 10 (ablation) | FedAvg + GCD and FedAvg + AGCL · CIFAR-10, CIFAR-100, and ImageNet-100 | All accuracy: 78.7 -> 63.4; 54.2 -> 47.3; 66.4 -> 62.3; 82.5 -> 68.1; 54.2 -> 52.2; 73.1 -> 67.5 | Hurt |
| Set the CSA sampling count to N S = 0 (cluster component means only) (ablation) | AGCL · Fed-GCD | All accuracy | Helped |
| Increase the CSA sampling count to more than 2 (ablation) | AGCL · Fed-GCD | All accuracy | Hurt |
Surprising findings
- Federated GCD suffers a large accuracy gap versus centralized training, especially on fine-grained datasets.
Limitations
- Server still needs labeled data for class number estimation and semi-supervised FINCH.
- Performance gap remains relative to centralized training under heterogeneity.
- Only image classification datasets evaluated with a single ViT backbone.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-100 CUB-200 Stanford Cars Oxford-IIIT Pet
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-100 CUB-200 Stanford Cars Oxford-IIIT Pet
Metrics: clustering accuracy (All, Old, New)
Baselines: FedAvg + GCD FedAvg + PCL FedAvg + GPC FedAvg + GCL FedProx + AGCL Centralized-GCD Centralized-PCL Centralized-GCL
Builds on: FedAvg GCD (Generalized Category Discovery with self-supervised contrastive learning) prototypical contrastive learning FINCH GPC
Closest papers by meaning
- Towards Understanding Parametric Generalized Category Discovery on Graphs (ICML 2025 · similarity 0.89)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.88)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.88)
- When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach (CVPR 2025 · similarity 0.88)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · 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.