ELFS: Entropy-based Loss Function Selection for Global Model Accuracy in Federated Learning (built on FL-bench)

Official implementation of ELFS (Entropy-based Loss Function Selection) for Federated Learning (FL), including code, configs, and all paper experiments. Built on top of FL-bench for reproducibility.

TL;DR

• What: Dynamically selects the loss function (e.g., Cross-Entropy, Focal) per round/aggregation step based on entropy signals, improving global model accuracy under Non-IID client distributions.
• Why: Different loss functions excel under different data skews/hardness; ELFS chooses the right one when it matters.
• How: Compute entropy statistics → compare to threshold/schedule → pick the loss from a candidate pool → train and aggregate.

Environment

• Python 3.8+
• PyTorch (CUDA optional but recommended)
• Install dependencies:

python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Adjust PyTorch/CUDA to your system if needed.

Quickstart

# 1) Generate Non-IID data 
python generate_data.py -d mnist -a 0.1 -cn 100

# 2) Train (elfs)
python main.py method=fedavg

That’s it—these two commands reproduce the basic experiment.

Acknowledgement

This work builds on FL-bench: https://github.com/KarhouTam/FL-bench

License

Choose a license for this repository (e.g., MIT/Apache-2.0) and ensure FL-bench’s license terms are followed and attributed.