Scholarlink

EEG signals hold great promise as biometric identifiers due to their invisibility and adaptability to high-security application scenarios. However, extracting reliable EEG identity features remains a challenge, primarily due to interference from device-related variations and the inherent differences in the subject’s state across multiple sessions. Current methods often treat each training session as a separate domain, which is problematic due to the differing data distributions across sessions. While many multi-source domain adaptation techniques attempt to bridge the domain gap between multiple source and target domains individually, they fail to account for the interrelationships between domain-invariant features during distribution alignment. In this chapter, we propose a novel multi-source domain adaptation framework, the Tensorized Spatial-frequency Attention Network (TSFAN), designed to enhance the performance of EEG-based brain fingerprint identification in the target domain. Specifically, TSFAN models the significant relationships between domain-invariant features using a tensorized attention mechanism. This mechanism effectively incorporates appropriate spatial-frequency representations from both pairwise source-target domains as well as cross-source domains, all while mitigating the impact of distribution discrepancies among the source domains. To address the issue of dimensionality, TSFAN is further approximated in the Tucker format. By leveraging the low-rank properties of the Tucker decomposition, TSFAN is able to scale linearly with the number of domains, offering significant flexibility for extension to scenarios involving any number of sessions. Extensive experiments conducted on representative benchmark datasets demonstrate that TSFAN outperforms state-of-the-art methods, achieving superior results in terms of identification accuracy. Furthermore, our electrode selection analysis reveals that brainprint features across different sessions are distributed across various brain regions. Notably, a selection of 20 electrodes, based on the standard 10–20 system, proves to be sufficient for extracting stable and reliable identity information.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-task and Multi-session Brain Fingerprint Identification with Attention Neural Network with Domain Adaptation Learning

  • 章节

  • 出版模式：

    开放获取

  • 发表日期：

    2025年

查看全文
  • Wanzeng Kong,
  • Xuanyu Jin

摘要

EEG signals hold great promise as biometric identifiers due to their invisibility and adaptability to high-security application scenarios. However, extracting reliable EEG identity features remains a challenge, primarily due to interference from device-related variations and the inherent differences in the subject’s state across multiple sessions. Current methods often treat each training session as a separate domain, which is problematic due to the differing data distributions across sessions. While many multi-source domain adaptation techniques attempt to bridge the domain gap between multiple source and target domains individually, they fail to account for the interrelationships between domain-invariant features during distribution alignment. In this chapter, we propose a novel multi-source domain adaptation framework, the Tensorized Spatial-frequency Attention Network (TSFAN), designed to enhance the performance of EEG-based brain fingerprint identification in the target domain. Specifically, TSFAN models the significant relationships between domain-invariant features using a tensorized attention mechanism. This mechanism effectively incorporates appropriate spatial-frequency representations from both pairwise source-target domains as well as cross-source domains, all while mitigating the impact of distribution discrepancies among the source domains. To address the issue of dimensionality, TSFAN is further approximated in the Tucker format. By leveraging the low-rank properties of the Tucker decomposition, TSFAN is able to scale linearly with the number of domains, offering significant flexibility for extension to scenarios involving any number of sessions. Extensive experiments conducted on representative benchmark datasets demonstrate that TSFAN outperforms state-of-the-art methods, achieving superior results in terms of identification accuracy. Furthermore, our electrode selection analysis reveals that brainprint features across different sessions are distributed across various brain regions. Notably, a selection of 20 electrodes, based on the standard 10–20 system, proves to be sufficient for extracting stable and reliable identity information.