Liu Honggang, Jin Xuanyu, Liu Dongjun, Kong Wanzeng, Tang Jiajia, Peng Yong
School of Computer Science, Hangzhou Dianzi University, Hangzhou, 310018 Zhejiang China.
Zhejiang Key Laboratory of Brain-Machine Collaborative Intelligence, Hangzhou, 310018 Zhejiang China.
Cogn Neurodyn. 2024 Oct;18(5):2897-2908. doi: 10.1007/s11571-024-10132-x. Epub 2024 May 29.
The electroencephalogram (EEG) signal is being investigated as a more confidential biometric for person identification. Despite recent advancements, a persistent challenge lies in the influence of variations in affective states. Affective states consistently exist during data collection, regardless of the protocol used. Additionally, the inherently non-stationary nature of EEG makes it susceptible to fluctuations in affective states over time. Therefore, it would be highly crucial to perform precise EEG-based person identification under varying affective states. This paper employed an integrated Multi-scale Convolution and Graph Pooling network (MCGP) to mitigate the impact of affective state variations. MCGP utilized multiple 1D convolutions at different scales to dynamically extract and fuse features. Additionally, a graph pooling layer with an attention mechanism was incorporated to generate hierarchical graph embeddings. These embeddings were concatenated as inputs for a fully connected classification layer. Experiments were conducted on the SEED and SEED-V dataset, revealing that MCGP achieved an average accuracy of 85.51% for SEED and 88.69% for SEED-V in cross-session conditions involving mixed affective states. Under single affective state cross-session scenario, MCGP achieved an average accuracy of 85.75% for SEED and 88.06% for SEED-V for the same affective states, while obtaining 79.57% for SEED and 84.52% for SEED-V for different affective states. Results indicated that, compared to the baseline methods, MCGP effectively mitigated the impact of variations in affective states across different sessions. In single affective state cross-session scenario, identification performance for the same affective states was slightly higher than that for different affective states.
脑电图（EEG）信号正作为一种更具保密性的生物特征用于身份识别而受到研究。尽管最近取得了进展，但一个持续存在的挑战在于情感状态变化的影响。在数据收集过程中，无论使用何种协议，情感状态始终存在。此外，EEG固有的非平稳特性使其容易随时间受到情感状态波动的影响。因此，在不同情感状态下进行基于EEG的精确身份识别至关重要。本文采用了一种集成的多尺度卷积和图池化网络（MCGP）来减轻情感状态变化的影响。MCGP利用不同尺度的多个一维卷积来动态提取和融合特征。此外，还引入了一个带有注意力机制的图池化层来生成分层图嵌入。这些嵌入被连接起来作为全连接分类层的输入。在SEED和SEED-V数据集上进行了实验，结果表明，在涉及混合情感状态的跨会话条件下，MCGP在SEED上的平均准确率为85.51%，在SEED-V上为88.69%。在单一情感状态跨会话场景中，对于相同情感状态，MCGP在SEED上的平均准确率为85.75%，在SEED-V上为88.06%；对于不同情感状态，在SEED上为79.57%，在SEED-V上为84.52%。结果表明，与基线方法相比，MCGP有效地减轻了不同会话中情感状态变化的影响。在单一情感状态跨会话场景中，相同情感状态下的识别性能略高于不同情感状态下的识别性能。
