Federated Learning (FL) enables collaborative training across distributed data silos without sharing raw data, but most deployments rely on centralized parameter-server architectures that create communication bottlenecks and limit scalability. Decentralized federated learning (DFL) mitigates this through peer-to-peer communication; however, existing approaches often employ stochastic or loosely defined communication schedules, making systematic analysis under heterogeneous environments difficult. We present DeTFL, a decentralized federated learning framework that introduces deterministic topology scheduling through pre-defined, round-synchronous communication graphs, enabling reproducible and fine-grained evaluation of topology-induced effects on convergence, communication efficiency, and system performance. Beyond compute-centric designs, DeTFL supports heterogeneous node roles, including computation, aggregation, and relay responsibilities, decoupling training logic from communication infrastructure while maintaining consistent execution across both simulation and real-world heterogeneous deployments. We further develop a convergence model capturing the interplay between topology-aware aggregation and system heterogeneity, providing theoretical insight into optimization dynamics and communication-induced variance. Experiments across multiple benchmarks, communication topologies, and heterogeneous hardware demonstrate that deterministic topology scheduling significantly reduces communication overhead while maintaining competitive accuracy, highlighting the critical role of communication structure in scalable federated learning.
In this tutorial, we will learn how to use the DeTFL framework to train/benchmark the FADNet model in the DeTFL settings with PyTorch.
Firstly, please install required dependencies. It is recommended to create a virtual environment. You also install libraries you need. In this tutorial, we provide a requirements.txt file, which contains necessary libraries to train the FADNet model.
PYTHON 3.9.5
pip install -r requirements.txt
pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu129We provide prepared autonomous driving datasets, please download and put them in the benchmark/FADNet/data folder.
After that, please download topology networks graph and extract them to the benchmark/FADNet/graph_utils folder.
<benchmark>
├── <FADNet>
│ ├── graph_utils
│ └── results
│ ├── amazon_us
│ ├── exodus
│ └── gaia
│ └── data
│ ├── driving_carla
│ ├── gaia
│ ├── amazon_us
│ ├── exodus
│ └── original
│ ├── driving_gazebo
│ ├── gaia
│ ├── amazon_us
│ ├── exodus
│ └── original
│ └── driving_udacity
│ ├── gaia
│ ├── amazon_us
│ └── exodus
To train our DeTFL framework on 3 autonomous driving datasets (Carla, Gazebo, and Udacity) with GAIA network, run:
bash train_scripts.shThe following table summarizes the benchmark configurations and resources for each experimental setup in DeTFL.
| No. | Dataset | Network | Topology | Checkpoints | Logs |
|---|---|---|---|---|---|
| 1 | Carla | GAIA | Ring | Download | Logs |
| 2 | Gazebo | GAIA | Ring | Download | Logs |
| 3 | Udacity | GAIA | Ring | Download | Logs |
This benchmark setup enables systematic evaluation of how communication topology and system heterogeneity influence training efficiency and convergence behavior.
📄 Related Paper
This repository accompanies the material report:
"Deterministic Topology Scheduling for Heterogeneous Decentralized Federated Learning".
The report provides theoretical analysis, convergence modeling, and extensive system-level evaluation of topology-aware decentralized federated learning.