EcoFL v0.1.0 — Initial Research Release

EcoFL v0.1.0 — Initial Research Release

This is the initial public research release of EcoFL (Energy-Conscious Federated Learning), a reproducible benchmarking framework for evaluating energy-aware federated learning on resource-constrained IoT edge devices.

EcoFL accompanies the manuscript:

“Can Federated Learning Go Green? EcoFL: A System-Level Energy-Aware Benchmark for IoT Edge Intelligence”

Main features

• Modular benchmarking framework for lightweight federated learning experiments.
• Support for three learning configurations:
  • centralized training,
  • standard federated learning using fixed-round FedAvg,
  • EcoFL energy-aware federated learning with adaptive round termination.
• Implementation of five lightweight model families:
  • Logistic Regression,
  • Random Forest,
  • XGBoost,
  • Multilayer Perceptron,
  • Isolation Forest.
• Synthetic IoT anomaly detection dataset generator with:
  • 50,000 samples,
  • 12 telemetry-inspired features,
  • configurable anomaly rate,
  • label noise,
  • Dirichlet-based non-IID client partitioning.
• TDP-based computation-side energy estimation using CPU utilization and training duration.
• System-level profiling of:
  • estimated training energy,
  • communication overhead,
  • inference latency,
  • process-level memory usage,
  • federated round counts.
• Reproducible experiments across five independent random seeds.
• Scheduler ablation study for convergence tolerance, patience, and CPU threshold.
• Statistical testing utilities based on Wilcoxon signed-rank tests and bootstrap confidence intervals.
• Figure and result generation scripts for reproducing the manuscript tables and plots.

Repository contents

This release includes:

• ecofl/ — core Python package,
• experiments/ — scripts for running the main benchmark, ablation study, statistical tests, and visualizations,
• results/ — aggregated results and experiment outputs,
• tests/ — minimal validation tests,
• README.md — installation and reproduction instructions,
• requirements.txt and environment.yml — environment specifications,
• pyproject.toml — package metadata,
• CITATION.cff — citation metadata,
• .zenodo.json — Zenodo metadata,
• reproduce.sh — convenience script for reproducing the benchmark workflow.

Reproducibility

The main experimental workflow can be reproduced with:

pip install -r requirements.txt
bash reproduce.sh

or step by step:

python experiments/run_experiments.py
python experiments/statistical_tests.py
python experiments/ablation_study.py
python experiments/visualize_results.py

The default configuration uses five independent seeds: 42, 43, 44, 45, 46.

Notes

This release is intended as a reproducible research companion for the EcoFL manuscript. The reported energy values are based on a TDP-based software profiling model and should be interpreted as computation-side energy estimates rather than direct physical power measurements. Future versions will extend EcoFL with physical hardware validation, wireless communication energy models, and additional real-world IoT datasets.

License

Released under the MIT License.

Recommended citation

Please cite this Zenodo release and the accompanying manuscript when using EcoFL in scientific work.