Research map / Federated learning and privacy / Communication-efficient FL
Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees
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 DSpodFL, a decentralized federated learning framework where per-iteration local SGD steps and pairwise model exchanges are random indicator variables, unifying DGD, DFedAvg and randomized gossip, with convergence guarantees for convex and non-convex losses.
Problem
Decentralized federated learning typically assumes a fixed number of local updates between model exchanges, ignoring heterogeneous and time-varying computation and communication resources across clients.
Approach
Model both local SGD occurrence at each client and pairwise model exchanges as arbitrary (time-varying, heterogeneous) Bernoulli indicator variables in a single update rule, so that DGD, DFedAvg and randomized gossip arise as special cases, and derive convergence bounds under this sporadicity.
Key results
- DSpodFL reaches target accuracies with significantly smaller delay than DGD, RG, sporadic SGD and DFedAvg baselines on FMNIST and CIFAR10 in IID and non-IID settings.
- Reported accuracy margins of 10-20% early in training and 10-40% improvement at a given delay for non-IID settings.
- Convergence: geometric rate for constant step size with an optimality gap, and O(ln K / sqrt K) sublinear rate with diminishing step size.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use sporadic local SGD and sporadic client-to-client aggregation together | DSpodFL vs. decentralized baselines · FMNIST and CIFAR10, IID | test accuracy at a given delay: 10-20% | Helped |
| Use sporadic local SGD and sporadic client-to-client aggregation together | DSpodFL vs. decentralized baselines · FMNIST and CIFAR10, non-IID | test accuracy at a given delay: 10-40% | Helped |
| Use sporadic local SGD and sporadic client-to-client aggregation together | DSpodFL vs. decentralized baselines · FMNIST, varying data distribution, graph radius, client count, and resource heterogeneity | test accuracy at a given delay | Helped |
| Use sporadic local SGD and sporadic client-to-client aggregation together | DSpodFL vs. decentralized baselines · FMNIST, 50 clients and heterogeneous bimodal resource probabilities | test accuracy at a given delay: 15% | Helped |
| Increase the number of labels per client, moving from non-IID toward IID | All methods · FMNIST | test accuracy at a given delay | Helped |
| Increase the random geometric graph radius | All baselines · FMNIST | test accuracy at a given delay | Helped |
Experimental setup
Datasets: Fashion-MNIST CIFAR10
Benchmarks: Fashion-MNIST CIFAR10
Metrics: test accuracy vs. total delay
Baselines: DGD Randomized gossip (RG) Sporadic SGDs DFedAvg
Builds on: decentralized gradient descent DFedAvg randomized gossip
Closest papers by meaning
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.91)
- NTK-DFL: Enhancing Decentralized Federated Learning in Heterogeneous Settings via Neural Tangent Kernel (ICML 2025 · similarity 0.91)
- Decentralized Directed Collaboration for Personalized Federated Learning (CVPR 2024 · similarity 0.90)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.90)
- Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.90)
- Federated Optimization with Doubly Regularized Drift Correction (ICML 2024 · similarity 0.89)
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.