Privacy-Preserving Attention-Weighted Multi-Source Domain Adaptation for EEG Motor Imagery

Abstract

Motor Imagery (MI) is essential in the Brain-Computer Interface (BCI). Given the multi-source nature of EEG instability across subjects/sessions and the privacy concerns when dealing with data from other institutions, the importance of multi-source-free domain adaptation (MSFDA) becomes evident in conducting subject-independent applications. However, current MI approaches are lack of MSFDA for considering different sources’ importance, which may lead to negative transfer. In this work, we propose a novel two-stage MSFDA framework, namely Privacy-preserving Attention-Weighted domain adaptation (PAW), which is the first to emphasize different source importance in EEG MI by identifying crucial features across subjects and sessions. A domain discriminator is designed to find robust features within sessions in the training phase. In the adaptation phase, an attention weighted module determines the source importance, and a balanced confident set policy mitigates noise and imbalance samples. PAW delivers excellent performance via empirical evaluations on several public datasets. In particular, its adaptation phase is applicable across MSFDA scenarios, highlighting its broad application values.