EEG_DLG: Reproducing “User Identity Protection in EEG-Based Brain–Computer Interfaces”

This repository is intended to support the reproducibility of the experiments reported in “User Identity Protection in EEG-Based Brain–Computer Interfaces” (hereafter referred to as the UIP-EEG paper).

The primary objective is to provide an organized workspace for running and documenting experiments related to identity protection in EEG-based brain–computer interfaces (BCIs), including (but not limited to) privacy-oriented training/evaluation settings and the management of intermediate artifacts produced during experimentation.

Scope

The reproduction effort focuses on:

• EEG motor imagery (MI) classification under subject-wise evaluation protocols
• User identity protection objectives and threat models discussed in the UIP-EEG paper
• Experiment traceability, including logs, cached artifacts, and result outputs

Repository Layout and Data Management

To ensure the repository remains lightweight and suitable for distribution, raw EEG datasets and large intermediate files are intentionally excluded from version control.

• data/: reserved for raw/downloaded datasets and processed caches
  • Contents are ignored by default via .gitignore.
  • Use .gitkeep files if you need to preserve directory structure in Git.
• output/: reserved for experimental outputs (e.g., checkpoints, metrics, figures)
  • Contents are ignored by default via .gitignore.
• run_scripts/: reserved for experiment launch scripts
  • *.log files are ignored by default.

Dataset Download

Use the unified downloader to fetch any subset of the supported datasets:

python3 data/download_datasets.py --which MI1,MI2,P300
python3 data/download_datasets.py --which MI2_IDremovalNet
python3 data/download_datasets.py --which MI1
python3 data/download_datasets.py --which MI2,P300

By default, P300 raw archives are downloaded to data/P300/raw and then preprocessed into the .mat files used by the training scripts. To only fetch the raw P300 archives, add --skip_p300_preprocess.

Reproducibility Recommendations

For rigorous reproducibility consistent with academic practice:

• Record software versions (Python, PyTorch, CUDA/cuDNN) and hardware (GPU model).
• Fix random seeds for Python/NumPy/PyTorch.
• Archive the exact configuration of hyperparameters and data splits used to generate each result.

Reference

If you use this repository in a scientific context, please cite the UIP-EEG paper:

“User Identity Protection in EEG-Based Brain–Computer Interfaces.”

eg for train_user_only.py
python -m scripts.train_user_only --epochs 600 --batch_size 8 --lr 2e-3 --user_hidden_dim 128 --user_dropout 0.5 --normalize channel --weight_decay 1e-4

eg for train_eegnet_MI1.py
python -m scripts.train --task_epochs 100 --user_epochs 100 --batch_size 8 --task_lr 2e-3 --user_lr 1e-3 --weight_decay 1e-4 --seeds 0 --dataset MI1 --model EEGNet