LFighter: Defending against Label-Flipping Attacks in Federated Learning

Official PyTorch implementation of the paper "LFighter: Defending against the Label-Flipping Attack in Federated Learning", published in Neural Networks, vol. 170, pp. 111–126, Elsevier, 2024.

---

Overview

LFighter is a robust aggregation defense mechanism designed to protect federated learning systems against label-flipping attacks — a class of data poisoning attacks in which malicious participants (clients) deliberately mislabel training samples to corrupt the global model's behavior on specific target classes.

This repository provides a fully reproducible implementation of all experiments reported in the paper across three benchmark datasets and multiple model architectures, under both IID and non-IID data distributions with up to 40% malicious participants.

---

Paper

LFighter: Defending against the Label-Flipping Attack in Federated Learning Najeeb M. Jebreel, Josep Domingo-Ferrer, David Sánchez, Alberto Blanco-Justicia Neural Networks, vol. 170, pp. 111–126, Elsevier, 2024 🔗 Read on ScienceDirect

---

Repository Structure

LFighter/
├── notebooks/                      # Jupyter notebooks for each benchmark
│   ├── Experiments_MNIST.ipynb
│   ├── Experiments_CIFAR10.ipynb
│   └── Experiments_IMDB_IID.ipynb
├── src/                            # Core Python source modules
│   ├── aggregation.py              # Aggregation rules (LFighter + baselines)
│   ├── datasets.py                 # Dataset loaders
│   ├── environment_federated.py    # Federated learning environment
│   ├── experiment_federated.py     # Experiment runner
│   ├── models.py                   # Model architectures
│   ├── sampling.py                 # Data sampling utilities
│   └── utils.py                    # General utilities
├── data/                           # Place IMDB dataset here (see below)
├── checkpoints/                    # Model checkpoints (auto-generated)
├── results/                        # Experiment results (auto-generated)
├── figures/                        # Result figures referenced in this README
│   ├── main_results.PNG
│   └── stability_all.PNG
├── requirements.txt
└── README.md

---

Installation

Prerequisites

Dependency	Version
Python	≥ 3.6
PyTorch	≥ 1.6
TensorFlow	≥ 2.0

Setup

git clone https://github.com/najeebjebreel/LFighter.git
cd LFighter
pip install -r requirements.txt

---

Datasets

Dataset	Access	Notes
MNIST	Automatic	Downloaded via PyTorch/TensorFlow data loaders
CIFAR-10	Automatic	Downloaded via PyTorch/TensorFlow data loaders
IMDB	Manual	See instructions below

IMDB Manual Setup

1. Download the preprocessed dataset from Google Drive.
2. Save the file as imdb.csv in the following path:

LFighter/data/imdb.csv

---

Reproducing Experiments

Open the Jupyter notebook corresponding to the benchmark of interest and follow the inline instructions:

jupyter notebook notebooks/Experiments_MNIST.ipynb
jupyter notebook notebooks/Experiments_CIFAR10.ipynb
jupyter notebook notebooks/Experiments_IMDB_IID.ipynb

Each notebook walks through data loading, federated training, attack simulation, and defense evaluation with inline instructions.

---

Results

Attack Robustness

The table below reports LFighter's classification robustness under a label-flipping attack with 40% malicious participants, compared against baseline aggregation strategies.

Source Class Accuracy Stability

The figure below shows the per-round source class accuracy under the label-flipping attack (40% attackers) for the CIFAR-10/ResNet-18/non-IID and IMDB/BiLSTM settings, illustrating LFighter's training stability relative to undefended baselines.

---

Citation

If you use this code or build upon this work, please cite:

@article{jebreel2024lfighter,
  title     = {LFighter: Defending against the label-flipping attack in federated learning},
  author    = {Jebreel, Najeeb Moharram and Domingo-Ferrer, Josep and S{\'a}nchez, David and Blanco-Justicia, Alberto},
  journal   = {Neural Networks},
  volume    = {170},
  pages     = {111--126},
  year      = {2024},
  publisher = {Elsevier},
  doi       = {10.1016/j.neunet.2023.11.047}
}

---

Acknowledgment

This research was funded by the European Commission (projects H2020-871042 "SoBigData++" and H2020-101006879 "MobiDataLab"), the Government of Catalonia (ICREA Acadèmia Prizes to J. Domingo-Ferrer and D. Sánchez, grant no. 2021 SGR 00115, and FI_B00760 grant to N. Jebreel), and MCIN/AEI/10.13039/501100011033 and "ERDF A way of making Europe" under grant PID2021-123637NB-I00 "CURLING". The authors are with the UNESCO Chair in Data Privacy, but the views in this paper are their own and are not necessarily shared by UNESCO.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Affiliation

Developed at the CRISES Research Group, Universitat Rovira i Virgili (URV), Tarragona, Catalonia.

---

Contact

For questions or issues, please open a GitHub issue or contact Najeeb Jebreel.