Adaptive Class-wise Multicentric Prototype Source-Free Domain Adaptation for Privacy-Preserving BCIs

Recently, conventional domain adaptation (DA) methods have demonstrated promising performance in cross-subject classification in electroencephalogram (EEG)-based brain-computer interfaces (BCIs). However, these methods require direct access to labeled source subject data, potentially compromising biometric privacy. Source-free domain adaptation (SFDA) addresses this issue by leveraging pre-trained source models with prototype-based pseudo-labeling, thereby eliminating the need for source data access. Nevertheless, current SFDA methods rely on oversimplified representations that fail to adequately capture EEG dynamics, resulting in two critical drawbacks: (1) oversimplified representations, using single centroids, fail to adequately capture complex neural manifolds, leading to intra-class collapse; (2) inadequate feature discrimination leads to error propagation in pseudo-labeling. To overcome these challenges, we propose the Adaptive Class-wise Multicentric Prototype-based SFDA (ACMP-SFDA) framework, which improves performance in new subjects while safeguarding personal privacy. Specifically, ACMP-SFDA dynamically constructs multiple prototypes per class to capture non-stationary EEG dynamics. Besides, we integrate semantic contrastive learning to improve inter-class discriminability while preserving the intrinsic structure of intra-class neural manifolds. Extensive experiments conducted across two BCI paradigms (motor imagery and affective BCI) demonstrate that ACMP-SFDA outperforms state-of-the-art methods, achieving 1.36$\%$, 1.96$\%$, and 1.94$\%$ accuracy improvements on MI2014001, MI2015001, and SEED, respectively, in cross-subject tasks.