EEG feature filter and disguising

This is our code for Neurocomputing paper "Information-preserving Feature Filter for Short-term EEG signals". This feature filter can filter out personal identities with desired features kept. The designed feature filter can be used for privacy protection. Our papers (feature filter, identity disguising model) are released online.

Data preprocessing

For the data preprocessing step, simply run the code listed below in order.

1. data splitting
   • within-subject data splitting: \DataPreprocessing\data_split_within.py
   • cross-subject data spplitting: \DataPreprocessing\data_split_cross.py
2. EEG signals to images
   • \DataPreprocessing\eegtoimg.py

Generate EEG images with dummy identities (Domain Y)

please run the following code in order:

1. time-frequency conversion:
   • \DataPreprocessing\t2f.py
   • file generated:
2. From EEG spectrums to EEG images with dummy identities:
   • \DataPreprocessing\grand_avg.py
3. joint training set:
   • \DataPreprocessing\gen_joint_train_set.py
4. the following .mat file generated before should be placed in the folder \EEG_idendity_disguising\datasets\eeg
   • eeg_dummy_images_w_label_step3_within.mat
   • eeg_images_train_augmented_within.mat
   • uci_eeg_images_test_within.mat
   • uci_eeg_images_train_within.mat
   • uci_eeg_images_validation_within.mat

Train the EEG disguising Model (feature filter)

• First use data_extra_combined_label.ipynb in the folder \EEG_idendity_disguising\datasets\eeg to generate training set with extra combined label
• Run the script:

python -m visdom.server

• Train the model with the semantic constraint on alocoholism feature

python train.py --name cycada_alcoholism --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model cycle_gan_semantic --lambda_A 1 --lambda_B 1 --lambda_identity 0 --no_flip --batchSize 64 --dataset_mode EEG --dataroot ./datasets/eeg/ --data_real uci_eeg_images_train_within_extra.mat --data_dummy eeg_dummy_images_w_label_step3_within_extra.mat --which_direction BtoA --feature alcoholism --num_classes 2

• Train the model with the semantic constraint on stimulus

python train.py --name cycada_stimulus --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model cycle_gan_semantic --lambda_A 1 --lambda_B 1 --lambda_identity 0 --no_flip --batchSize 64 --dataset_mode EEG --dataroot ./datasets/eeg/ --data_real uci_eeg_images_train_within_extra.mat --data_dummy eeg_dummy_images_w_label_step3_within_extra.mat --which_direction BtoA --feature stimulus --num_classes 5

• Train the model with the semantic constraint on the combined label (alcoholism + stimulus)

python train.py --name cycada_combined --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model cycle_gan_semantic --lambda_A 1 --lambda_B 1 --lambda_identity 0 --no_flip --batchSize 64 --dataset_mode EEG --dataroot ./datasets/eeg/ --data_real uci_eeg_images_train_within_extra.mat --data_dummy eeg_dummy_images_w_label_step3_within_extra.mat --which_direction BtoA --feature combined --num_classes 10

• Train the model without semantic constraint

python train.py --name cyclegan --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model cycle_gan --lambda_A 1 --lambda_B 1 --lambda_identity 0 --no_flip --batchSize 64 --dataset_mode EEG --dataroot ./datasets/eeg/ --data_real uci_eeg_images_train_within_extra.mat --data_dummy eeg_dummy_images_w_label_step3_within_extra.mat --which_direction BtoA

Train the Classification Model

• in the directory /EEG_identity_disguising/validation
• train the ResNet classifier
  • Select the ResNet model (ResNet18|ResNet34|ResNet50) and the classification task (alcoholism|stimulus|id) when training, for example:

python3 resnet_classification_model.py --model ResNet18 --feature alcoholism

The script above train a ResNet-18 model to perform the alcoholism classification task.

Test the EEG disguising Model

• several essential option in the script
  • --name checkpoints folder name
  • --classifier which classifier is used for evaluation
  • --test_all use this argument to require the validation on the model from all checkpoints saved every 5 training epochs
  • --which_epoch if test_all not specified, select one epoch of the model to test
• the example script below load checkpoints from the checkpoints cycada_alcoholism_v2, and requires to test the model from all training epochs

python validation.py --name cycada_alcoholism_v2 --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model test --no_flip --batchSize 32 --dataset_mode EEGsingle  --dataroot ./datasets/eeg/ --data uci_eeg_images_validation_within.mat  --which_direction BtoA --phase train --how_many 100000 --classifier ResNet34 --test_all

• the scripts below are the ones with the experiemnt results I put in the report.

python validation.py --name cycada_alcoholism_v2 --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model test --no_flip --batchSize 32 --dataset_mode EEGsingle  --dataroot ./datasets/eeg/ --data uci_eeg_images_validation_within.mat  --which_direction BtoA --phase train --how_many 100000 --classifier ResNet34 --which_epoch 190

python validation.py --name cycada_stimulus_v2 --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model test --no_flip --batchSize 32 --dataset_mode EEGsingle  --dataroot ./datasets/eeg/ --data uci_eeg_images_validation_within.mat  --which_direction BtoA --phase train --how_many 100000 --classifier ResNet34 --which_epoch 45

python validation.py --name cycada_combined_v2 --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model test --no_flip --batchSize 32 --dataset_mode EEGsingle  --dataroot ./datasets/eeg/ --data uci_eeg_images_validation_within.mat  --which_direction BtoA --phase train --how_many 100000 --classifier ResNet34 --which_epoch 150

python validation.py --name cyclegan_v2 --resize_or_crop=None --loadSize=32 --fineSize=32 --which_model_netD n_layers --n_layers_D 3 --model test --no_flip --batchSize 32 --dataset_mode EEGsingle  --dataroot ./datasets/eeg/ --data uci_eeg_images_validation_within.mat  --which_direction BtoA --phase train --how_many 100000 --classifier ResNet34 --which_epoch 40

The full version of the project with code, datasets and checkpoints are uploaded to this share link: code, datasets and checkpoints

*You may need to downgrade scipy to 1.1.0

pip install scipy==1.1.0

if you find this code useful, please kindly cite

@article{yao2020information,
  title={Information-preserving Feature Filter for Short-term EEG signals},
  author={Yao, Yue and Plested, Josephine and Gedeon, Tom},
  journal={Neurocomputing},
  year={2020},
  publisher={Elsevier}
}