Research map / Continual learning / Domain and prompt-based
C$^2$Prompt: Class-aware Client Knowledge Interaction for 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 C²Prompt, an exemplar-free prompt-based federated continual learning method that improves class-wise knowledge coherence across clients via a local class distribution compensation prompt and class-aware weighted prompt aggregation.
Problem
Federated continual learning suffers from both temporal and spatial forgetting; existing prompt-based FCL methods ignore class-wise knowledge coherence during inter-client prompt communication, causing intra-class distribution gaps and inter-class prompt confusion.
Approach
Estimate global per-class Gaussian distributions from client statistics and train compensation prompts to align local features to them (LCDC), and weight prompt aggregation on the server by class-affinity histograms collected from clients (CPA), built on CODAPrompt with Powder's distillation loss.
Key results
- On ImageNet-R, C²Prompt reaches 87.20% Avg versus 84.69% for Powder (the strongest baseline), and on DomainNet 78.88% vs 75.98%.
- Under the FCIL setting on ImageNet-R (FAA), C²Prompt improves over LoRM by 2.75/5.91/2.29 points at beta=0.5/0.1/0.05.
- On CIFAR-100 it reports 97.32% Avg vs 95.78% for Powder; ablations show LCDC alone adds 1.88% and CPA alone adds 1.33% on ImageNet-R.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LCDC (global class distribution estimation and local class distribution compensation) to the baseline (ablation) | C²Prompt · ImageNet-R | Avg: 1.88% | Helped |
| Add CPA (class-aware prompt aggregation) to the baseline (ablation) | C²Prompt · ImageNet-R | Avg: 1.33% | Helped |
| Add both LCDC and CPA to the baseline (ablation) | C²Prompt · ImageNet-R | Avg: 2.51% | Helped |
| Use C²Prompt instead of Powder | FCL methods · ImageNet-R | Avg: 2.51% | Helped |
| Use C²Prompt instead of Powder | FCL methods · DomainNet | Avg: 2.90% | Helped |
| Use C²Prompt instead of Powder | FCL methods · CIFAR-100 | Avg: 1.54% | Helped |
Surprising findings
- Fed-CODAP and Powder show negative forgetting on ImageNet-R, which the authors attribute to new tasks facilitating old-task learning when samples are limited.
- C²Prompt's backward transfer is lower than Fed-CODAP and Powder on ImageNet-R and DomainNet, which the authors attribute to stronger new-task learning leaving less room for improvement on seen tasks.
Limitations
- Requires an initial round of class distribution communication, adding some communication cost.
- Client-specific prompts slightly increase computing and storage overhead relative to existing prompt-based methods.
- Backward transfer is weaker than Fed-CODAP and Powder on ImageNet-R and DomainNet.
Experimental setup
Datasets: ImageNet-R DomainNet CIFAR-100
Benchmarks: ImageNet-R DomainNet CIFAR-100
Metrics: Avg AIA FM FT BT CT FAA
Baselines: FedWEIT CFeD GLFC FedSpace Fed-L2P Fed-Dual Fed-CODAP Fed-CPrompt Powder PILoRA Fed-MOS LoRM
Builds on: CODAPrompt Powder L2P DualPrompt FedAvg
Closest papers by meaning
- Federated Continual Learning via Prompt-based Dual Knowledge Transfer (ICML 2024 · similarity 0.90)
- Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.90)
- Federated Class-Incremental Learning: A Hybrid Approach Using Latent Exemplars and Data-Free Techniques to Address Local and Global Forgetting (ICLR 2025 · similarity 0.88)
- Federated Continual Learning Goes Online: Uncertainty-Aware Memory Management for Vision Tasks and Beyond (ICLR 2025 · similarity 0.88)
- Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning (NeurIPS 2025 · similarity 0.87)
- CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class Incremental Learning (AAAI 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.