Research map / Federated learning and privacy / Personalized and heterogeneous FL
CAN: Leveraging Clients As Navigators for Generative Replay in Federated Continual 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 CAN, a client-centric generative replay method for federated continual learning that uses client experts to guide synthetic data generation and personalizes replay buffers per client based on forgetting profiles.
Problem
Generative replay in federated continual learning ignores FL-specific properties such as non-IID client data, where clients have varying expertise and forgetting patterns, leading to suboptimal synthetic data and uniform replay buffers.
Approach
Treats clients as domain experts: for each class the most accurate client guides a generator via a navigation loss, a reference model trained on synthetic data adds a gap-driven KL loss to capture subtle features, and each client's replay buffer is reallocated across tasks according to its measured forgetting weights.
Key results
- On CIFAR100 5-task IID, CAN reaches 48.16% average accuracy versus 45.92% for LANDER at the same 1,250 buffer size.
- On ImageNet100 NIID(0.1) over 5 tasks, CAN averages 28.18% accuracy versus 14.97% for LANDER.
- CAN outperforms LANDER at smaller buffer sizes (e.g., 500) on CIFAR100, with comparable results to LANDER at larger buffers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAN instead of LANDER | CIFAR100 continual learning · CIFAR100, 5-task and 10-task settings across IID and NIID(1), NIID(0.5), and NIID(0.1) | Average accuracy (A): 48.16 vs 45.92; 45.93 vs 44.51; 42.16 vs 41.63; 39.88 vs 37.94; 34.41 vs 31.29; 32.59 vs 31.07; 28.06 vs 26.93; 23.60 vs | Helped |
| Use CAN instead of LANDER | ImageNet100 continual learning · ImageNet100, NIID(0.5) and NIID(0.1), Tasks 2–5 | Average accuracy: 50.15 vs 40.40; 40.87 vs 31.30; 32.70 vs 23.92; 26.52 vs 17.48; 40.90 vs 26.20; 31.00 vs 15.90; 23.30 vs 10.75; 17.50 vs | Helped |
| Increase buffer size | CAN · TinyImageNet, 5 tasks, IID and non-IID settings | Accuracy | Mixed |
Surprising findings
- Accuracy on TinyImageNet shows a non-monotonic 'V' trend with buffer size for CAN, suggesting a trade-off between synthetic data quality and quantity.
Limitations
- Gains on TinyImageNet are non-monotonic with buffer size, so larger buffers do not reliably help CAN.
- Client expertise map relies on per-class accuracy of clients, and the paper does not detail the cost of selecting experts in large federations.
Experimental setup
Datasets: CIFAR100 TinyImageNet ImageNet100
Benchmarks: CIFAR100 TinyImageNet ImageNet100
Metrics: average accuracy average forgetting score
Baselines: Vanilla FedEWC FedLwF FedWeIT TARGET LANDER DDDR
Builds on: TARGET LANDER generative replay elastic weight consolidation learning without forgetting
Closest papers by meaning
- Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.92)
- Towards Efficient Replay in Federated Incremental Learning (CVPR 2024 · similarity 0.89)
- Federated Continual Learning Goes Online: Uncertainty-Aware Memory Management for Vision Tasks and Beyond (ICLR 2025 · similarity 0.88)
- HERO: A Heterogeneity-Aware Benchmark Library for Federated Continual Learning (NeurIPS 2026 · similarity 0.88)
- Resource-Constrained Federated Continual Learning: What Does Matter? (NeurIPS 2025 · similarity 0.87)
- Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning (NeurIPS 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.