chamwen/MSDT

Code for Multi-Source Decentralized Transfer for Privacy-Preserving BCIs

Multi-Source Decentralized Transfer for Privacy-Preserving BCIs

This repository contains codes of our paper https://ieeexplore.ieee.org/document/9894428

Datasets

The MI datasets can be downloaded and processed with MOABB in http://moabb.neurotechx.com/docs/datasets.html

Prerequisites:

• python == 3.7.6

• pyriemann == 0.2.6

• PyTorch == 1.8.0

• mne == 0.20.7

• numpy, scipy, sklearn

Running the code

Code files introduction:

utils/ -- necessary function files

source_train_multi_mi.py -- demo file, source models pre-training.

target_adapt_msdt_mi.py -- demo file, gray box MSDT.

target_adapt_msdt_kd.py -- demo file, black box MSDT.

Notes

The codes are only for reference. In the early version, in the model pre-training stage, we set the learning rate of the feature extractor to 1/10 of the feature extractor, which has been revised as the same learning rate. The cross-subject classification results with this version are similar to the paper.