Research map / Federated learning and privacy / Communication-efficient FL
Decentralized Directed Collaboration for Personalized Federated 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 DFedPGP, a decentralized personalized federated learning method that combines partial model personalization (private linear classifier) with stochastic gradient push over time-varying directed graphs, with an O(1/sqrt(T)) non-convex convergence guarantee.
Problem
Personalized federated learning in decentralized peer-to-peer settings suffers from suboptimal personalization and unguaranteed convergence when using undirected, symmetric topologies under data and resource heterogeneity.
Approach
Decouple the model into a shared feature extractor and a private linear classifier, and exchange only shared parameters over a directed, asymmetric topology using push-sum with de-biasing weights, with alternating local updates of the personal and shared parts.
Key results
- On CIFAR-10 Dirichlet α=0.3, DFedPGP reaches 85.61% test accuracy vs 84.50% for FedRep (best comparator).
- On CIFAR-100 Dirichlet α=0.3, DFedPGP reaches 66.26% at α=0.1 vs 60.23% for FedBABU; it reaches 45% accuracy in 111 rounds vs 187 for DFedAvgM on Dir-0.3.
- Under computation heterogeneity on CIFAR-10 Dir-0.3, DFedPGP attains 83.63% vs 83.08% for FedRep.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DFedPGP instead of FedRep | CIFAR-10 PFL · CIFAR-10 Dirichlet α=0.3 | test accuracy (%): 84.50±.05 -> 85.61±.05 | Helped |
| use DFedPGP instead of FedBABU | CIFAR-100 PFL · CIFAR-100 Dirichlet α=0.1 | test accuracy (%): 60.23±.07 -> 66.26±.25 | Helped |
| use DFedPGP instead of OSGP | CIFAR-100 PFL · CIFAR-100 Pathological c=10 | test accuracy (%): 71.08±.52 -> 72.19±.21 | Helped |
| add partial personalization to DFedAvgM (ablation) | DFedAvgM · CIFAR-10 Dirichlet α=0.3 | test accuracy (%): 82.49 -> 84.69 | Helped |
| add partial personalization to DFedAvgM (ablation) | DFedAvgM · CIFAR-10 Pathological c=2 | test accuracy (%): 90.23 -> 90.90 | Helped |
| add directed communication and partial personalization to OSGP (ablation) | OSGP · CIFAR-10 Dirichlet α=0.3 | test accuracy (%): 83.14 -> 85.61 | Helped |
| add directed communication and partial personalization to OSGP (ablation) | OSGP · CIFAR-10 Pathological c=2 | test accuracy (%): 90.72 -> 91.26 | Helped |
| increase the number of participating neighbors (ablation) | DFedPGP · CIFAR-10 Dirichlet α=0.3 | test accuracy | Helped |
| reduce participating clients to 5 or 10 rather than 50, 100, or 200 (ablation) | DFedPGP · CIFAR-10 Dirichlet α=0.3 | test accuracy | Helped |
| increase communication network connectivity and exchange with more neighbors | DFedPGP · non-convex convergence analysis | convergence rate | Helped |
Surprising findings
- Under extreme pathological heterogeneity, fewer classes per client yields higher accuracy, since the local task becomes easier.
Limitations
- Only random client selection is used; no adaptive neighbor selection.
- Experiments limited to image classification with ResNet-18 and group normalization.
- Theory assumes bounded variance and gradient diversity assumptions that are mild but idealized.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet
Metrics: test accuracy communication rounds to target accuracy
Baselines: Local FedAvg FedPer FedRep FedBABU Ditto DFedAvgM OSGP Dis-PFL
Builds on: Stochastic Gradient Push (SGP) FedRep FedPer FedBABU Push-sum
Closest papers by meaning
- Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees (ICLR 2025 · similarity 0.90)
- Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees (ICML 2025 · similarity 0.90)
- pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning (NeurIPS 2024 · similarity 0.89)
- FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.89)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.89)
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.88)
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.