BART-FL: A Backdoor Attack-Resilient Federated Aggregation Technique for Cross-Silo Applications

BART-FL (Backdoor-Aware Robust Training for Federated Learning) is a novel defense-oriented framework that enhances the robustness of Federated Learning (FL) against backdoor and poisoning attacks. It integrates Principal Component Analysis (PCA) and clustering-based filtering to isolate and suppress malicious client updates while maintaining accuracy on clean data. Designed for cross-device federated environments, BART-FL provides explainable and privacy-aware aggregation mechanisms to improve resilience against adversarial behavior.

📘 Read the paper on IEEE Xplore

🛠️ Requirements

• Python 3.9+
• PyTorch 2.0+
• Torchvision
• Pillow
• NumPy
• scikit-learn
• CUDA 11.8 (optional, for GPU support)

Install dependencies using:

pip install -r requirements.txt

📚 Datasets and Models

📚 Datasets

This project supports the following datasets for federated learning:

• CIFAR-10: Automatically downloaded via torchvision.datasets.CIFAR10.
• CIFAR-100: Automatically downloaded via torchvision.datasets.CIFAR100.
• LISA Traffic Light Dataset: Local dataset that must be placed under
  data/lisa/train/ and data/lisa/val/ folders.
  All images are resized to 32×32 resolution for model compatibility.

🤖 Models

This project supports the following neural network architectures for federated learning:

• VGG6: Lightweight convolutional model adapted for CIFAR and LISA datasets.
• AlexNet5: Simplified 5-layer AlexNet variant for efficient client-side training.
• ResNet18: Deep residual network with skip connections for robust performance.

🔧 Federated Learning with BART-FL

Train CIFAR-100 with ResNet18 and the Adaptive Patch Attack using the following command:

python main.py \
  --device GPU \
  --dataset CIFAR100 \
  --partition nonIID \
  --model_name ResNet18 \
  --aggregation bartfl \
  --num_rounds 100 \
  --n_clients 10 \
  --attack_type adaptive_patch \
  --n_attackers 4 \
  --poison_rate 0.7 \
  --alpha 0.2 \
  --target_class 0 \
  --batch_size 64 \
  --learning_rate 0.001 \
  --weight_decay 0.0001 \
  --trigger_dir ./triggers \
  --checkpoint_dir checkpoints \
  --output_dir result \
  --csv_filename result/bartfl-adaptive-ResNet18-cifar100.csv \
  --pca-dim 0.95 \
  --num_clusters 2 \
  --patience 5 \
  --random_seed 42 \
  --verbose

📜 Citation

If you use BART-FL in your research, please cite the following paper:

@ARTICLE{11172307,
  author={Mia, Md Jueal and Hadi Amini, M.},
  journal={IEEE Transactions on Machine Learning in Communications and Networking}, 
  title={BART-FL: A Backdoor Attack-Resilient Federated Aggregation Technique for Cross-Silo Applications}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={Training;Data models;Computational modeling;Filtering;Federated learning;Data privacy;Adaptation models;Transportation;Reflection;Principal component analysis;Federated Learning;Backdoor Attacks;Outlier Detection;Security;Cross-Silo Applications},
  doi={10.1109/TMLCN.2025.3611398}}

📘 Read the paper on IEEE Xplore: https://ieeexplore.ieee.org/document/11172307