Research map / Federated learning and privacy / Personalized and heterogeneous FL
Decentralized Dynamic Cooperation of Personalized Models 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 DCFCL, a decentralized federated continual learning framework where clients form dynamic, non-overlapping coalitions via a coalitional affinity game over overall (gradient + model) similarity, with merge-blocking and dynamic cooperative evolution algorithms to reach equilibrium per task phase.
Problem
Federated continual learning suffers from temporal catastrophic forgetting and spatial interference from aggregating heterogeneous client models into a global model; directly applying decentralized group aggregation to FCL yields ineffective groups when tasks change.
Approach
Clients form dynamic non-overlapping coalitions at each task phase: pairwise benefit is overall similarity (gradient cosine plus weighted model cosine), multi-client benefits are derived via a coalitional affinity game, and a merge-blocking algorithm finds a cooperative equilibrium that is re-evolved as tasks change.
Key results
- On EMNIST-LTP, DCFCL reaches 52.5% average accuracy vs 47.5% for the best baseline AF-FCL; on EMNIST-shuffle 78.3% vs 77.0% (AFCL).
- On CIFAR100 (4 tasks, 20 classes per task), DCFCL achieves 40.4% vs 37.4% for SCAFFOLD; on MNIST-SVHN-F it reaches 66.7%, below AF-FCL's 68.1%.
- Ablation on EMNIST-LTP: removing cooperative equilibrium (local training) drops accuracy to 12.3%, and removing knowledge distillation or overall similarity drops it to 50.3% and 45.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic cooperation (aggregation over coalitions) vs. local-only training (ablation) | decentralized FCL (DCFCL) · EMNIST-LTP | average accuracy: 12.3 to 52.5 | Helped |
| add knowledge distillation for classifier feature consistency (ablation) | DCFCL · EMNIST-LTP | average accuracy: 50.3 to 52.5 | Helped |
| add knowledge distillation for classifier feature consistency (ablation) | DCFCL · EMNIST-shuffle | average accuracy: 73.2 to 78.3 | Helped |
| add overall similarity (gradient plus model similarity) for benefit quantification (ablation) | DCFCL · EMNIST-LTP | average accuracy: 45.3 to 52.5 | Helped |
| add overall similarity (gradient plus model similarity) for benefit quantification (ablation) | DCFCL · EMNIST-shuffle | average accuracy: 73.7 to 78.3 | Helped |
| increase model-similarity weight epsilon from 0.2 to 1.0 or 0.4 variations (ablation) | DCFCL · EMNIST-shuffle | average accuracy: peak 80.2 at epsilon 0.4; 70.4 to 81.3 across values | Mixed |
| raise lambda (distillation weight) from 0.0 to 1.0 (ablation) | DCFCL · EMNIST-LTP | average accuracy: 50.3 to 55.7 at lambda 1.0; dips to 50.7 at 0.4 | Mixed |
| DCFCL outperforms FL/FCL baselines on CIFAR100 and EMNIST datasets but not on MNIST-SVHN-F | DCFCL · MNIST-SVHN-F | average accuracy: 66.7 vs AF-FCL 68.1 | Hurt |
Surprising findings
- On MNIST-SVHN-F, DCFCL (66.7%) does not beat the best baseline AF-FCL (68.1%), which the paper does not highlight.
- Local training without any aggregation collapses to 12.3% on EMNIST-LTP, far below FedAvg (32.5%).
Limitations
- Parameter sensitivity to lambda and epsilon, with adaptive mechanisms left for future work.
- Coalition formation is NP-hard in nature; large-scale deployment requires approximations that sacrifice optimality.
- Experiments are limited to small-scale image benchmarks with 8-10 clients.
Experimental setup
Datasets: EMNIST CIFAR100 MNIST SVHN FashionMNIST
Benchmarks: EMNIST-LTP EMNIST-shuffle CIFAR100 MNIST-SVHN-F
Metrics: average accuracy average forgetting
Baselines: FedAvg FedProx SCAFFOLD CFL Per-FedAvg PODNet+FedAvg PODNet+FedProx ACGAN+FedAvg ACGAN+FedProx FLwF2T FedCIL GLFC
Builds on: FedAvg CFL Per-FedAvg PODNet
Closest papers by meaning
- Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.89)
- Traceable Federated Continual Learning (CVPR 2024 · similarity 0.88)
- Federated Continual Learning via Prompt-based Dual Knowledge Transfer (ICML 2024 · similarity 0.88)
- 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 0.87)
- HERO: A Heterogeneity-Aware Benchmark Library for Federated Continual Learning (NeurIPS 2026 · 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.