Research map / Continual learning / Class-incremental
Class-wise Balancing Data Replay for Federated Class-Incremental 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 FedCBDR, a federated class-incremental learning method that builds class-balanced replay buffers using privacy-preserving global leverage-score sampling via SVD plus task-aware temperature scaling and re-weighted loss.
Problem
Data replay in federated class-incremental learning suffers from class imbalance in the replay buffer (due to lack of global awareness across clients) and between replayed and new classes, causing catastrophic forgetting.
Approach
Clients send randomly rotated (ISVD-encrypted) feature matrices to the server; the server performs SVD on the aggregated features and computes leverage scores to sample a globally balanced, importance-weighted replay buffer, then applies task-aware temperature scaling and sample re-weighting in the loss to counter old/new imbalance.
Key results
- FedCBDR reaches 64.11% on CIFAR10 (3 tasks, beta=0.5) vs 60.79% for the best baseline LANDER at beta=1.0 and 60.73% for Re-Fed at beta=1.0, with a 2%-15% Top-1 gain claimed over six baselines.
- On CIFAR100 (5 tasks, beta=1.0), LANDER (52.77%) slightly outperforms FedCBDR (52.06%), the only non-best result in the main table.
- Under 10 clients, FedCBDR reaches 59.80% on CIFAR10 (3 tasks, beta=0.5) vs 45.49% for Re-Fed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use global-perspective replay (GDR) instead of Finetune alone (ablation) | FedCBDR · CIFAR10, 3 tasks, β=0.5 | Top-1 accuracy: 38.71$_{\pm 3.7}$ -> 62.13$_{\pm 2.1}$ | Helped |
| Use global-perspective replay (GDR) instead of Finetune alone (ablation) | FedCBDR · CIFAR100, 5 tasks, β=0.5 | Top-1 accuracy: 16.75$_{\pm 2.6}$ -> 47.66$_{\pm 0.9}$ | Helped |
| Use global-perspective replay (GDR) instead of Finetune alone (ablation) | FedCBDR · TinyImageNet, 10 tasks, β=0.5 | Top-1 accuracy: 6.00$_{\pm 0.8}$ -> 17.89$_{\pm 0.5}$ | Helped |
| Use task-aware temperature scaling (TTS) instead of Finetune alone (ablation) | FedCBDR · CIFAR10, 3 tasks, β=0.5 | Top-1 accuracy: 38.71$_{\pm 3.7}$ -> 41.34$_{\pm 2.3}$ | Helped |
| Use task-aware temperature scaling (TTS) instead of Finetune alone (ablation) | FedCBDR · CIFAR100, 5 tasks, β=0.5 | Top-1 accuracy: 16.75$_{\pm 2.6}$ -> 17.14$_{\pm 0.4}$ | No clear effect |
| Use task-aware temperature scaling (TTS) instead of Finetune alone (ablation) | FedCBDR · TinyImageNet, 10 tasks, β=0.5 | Top-1 accuracy: 6.00$_{\pm 0.8}$ -> 6.92$_{\pm 0.3}$ | Helped |
| Add TTS to GDR (ablation) | FedCBDR · CIFAR100, 5 tasks, β=0.5 | Top-1 accuracy: 47.66$_{\pm 0.9}$ -> 49.76$_{\pm 2.7}$ | Helped |
| Increase the replay buffer size | FedCBDR · CIFAR10 | Top-1 accuracy: 51.99 -> 59.02 -> 63.81 | Helped |
| Increase the replay buffer size | FedCBDR · CIFAR100 | Top-1 accuracy: 40.12 -> 49.66 -> 55.94 | Helped |
| Choose a large new-task temperature (τ_new=2.0) | FedCBDR TTS · CIFAR100, β=0.5, 5-task split | Top-1 accuracy | Hurt |
Surprising findings
- LANDER, despite generating 10,240 synthetic samples per task, shows severe forgetting of early tasks (1.37% on CIFAR10 Task 1).
Limitations
- Feature transmission cost from clients to server is non-trivial and the authors plan to reduce it with lightweight sampling.
- Performance is sensitive to temperature and re-weighting hyperparameters (sensitivity analysis shows large variation).
- LANDER is best on one CIFAR100 setting (5 tasks, beta=1.0).
Experimental setup
Datasets: CIFAR10 CIFAR100 TinyImageNet
Benchmarks: CIFAR10 CIFAR100 TinyImageNet
Metrics: Top-1 accuracy
Baselines: Finetune FedEWC FedLwF TARGET LANDER Re-Fed
Builds on: Re-Fed LANDER GLFC
Closest papers by meaning
- Federated Class-Incremental Learning: A Hybrid Approach Using Latent Exemplars and Data-Free Techniques to Address Local and Global Forgetting (ICLR 2025 · similarity 0.90)
- Towards Efficient Replay in Federated Incremental Learning (CVPR 2024 · similarity 0.88)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.87)
- CAN: Leveraging Clients As Navigators for Generative Replay in Federated Continual Learning (ICML 2025 · similarity 0.87)
- C$^2$Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning (NeurIPS 2025 · similarity 0.86)
- Accurate Forgetting for Heterogeneous Federated Continual Learning (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.