<p>Due to their high temporal resolution, EEG signals have emerged as a primary modality for psychological state recognition in brain-computer interfaces. However, obtaining labeled EEG data can be challenging in certain scenarios. To overcome this limitation, some cross-domain adaptive methods have been proposed to perform recognition tasks by utilizing existing labeled data from other domains. During the process of domain adaptation, mismatches in sample distribution arise due to differences among domains in EEG signal acquisition, as well as the variability and non-smoothness of the EEG signal itself. Furthermore, accessing personal EEG signals from other domains is often impractical and problematic due to privacy concerns. To address the challenges posed by distribution differences and privacy protection, we propose a method called Source-free Adaptation EEGNet (SFA-EEGNet). Under the condition of source-free, the source subject’s data are unavailable. The model design considers both the spatial and temporal domains of EEG signals to extract features and utilizes pseudo-labeling to transfer knowledge from the source subject to the target. We construct a new dataset called Lie Detection and use the EEG public dataset SEED for the emotion recognition task. To demonstrate its versatility, we evaluate SFA-EEGNet in various adaptation scenarios. Experimental results indicate that SFA-EEGNet consistently achieves state-of-the-art performance when compared to multiple domain adaptation benchmarks.</p>