Due to significant inter-subject variability in feature distributions caused by the diversity of neural activity patterns, Electroencephalography (EEG)-based brain-computer interface (BCI) systems face considerable challenges in cross-subject EEG decoding. Though transfer learning has been widely introduced for knowledge transfer from source subject(s) to target subject and exhibited great success, a non-negligible issue is that source subjects' EEG data usually contains privacy information and should be protected. To address both issues, we propose a source-target bidirectional refined source-free domain adaptation (BR-SFDA) framework in this paper for privacy preserving cross-subject EEG classification. BR-SFDA makes improvements from two aspects under the popular 'pretraining and fine-tuning' paradigm. On one hand, it locally performs data augmentation and builds a multi-criteria fused metric to select representative EEG sample for model pre-training. On the other hand, a structured graph learning strategy is employed to better guide the model finetuning in a self-supervised manner. Both improvements collaborate respectively from the front-end and back-end, leading to a bidirectional refined SFDA framework. Extensive experiments are conducted on two tasks of cross-subject motor imagery decoding and emotion recognition, and the results on four datasets demonstrate that BR-SFDA achieves superior performance to some competitive models. Besides, the effectiveness of data augmentation and filtering, structured graph learning and domain adaptation is well evaluated.
Zhang et al. (Thu,) studied this question.