中国科学院心理研究所机构知识库

Lie detection techniques based on electroencephalogram (EEG) rely on the effective decoding of event-related potential (ERP). At present, manual design features are mainly used for EEG analysis. In recent years, the single-trial EEG classification method has made progress. End-to-end EEG classification methods can realize automatically extract features from EEG and classify, which lacks research and application in lie detection, also those methods cannot be directly applied in lie detection. In this study, we designed the autobiographical-based face recognition task based on a complex trial protocol (CTP) and the EEG of 18 subjects was collected. The application of different single-trial ERP classification methods in lie detection are studied. A class bootstrap method is proposed to solve the problem that the single-trial EEG decoding method cannot be applied to practice directly. The class bootstrap method was based on the assumption of data distribution, the probe stimulus was deduced by comparing the classification performance of classifiers that were trained when each category of stimulus images was set as probe stimuli. The experimental results show that the proposed class bootstrap method outperforms the traditional lie detection method and can accurately predict lies when only a small amount of EEG data is used.

摘要基于脑电(Electroencephalogram, EEG)的谎言检测技术依赖于对事件相关电位(Event-related potential ERP)的有效解码当前主要采用手工设计特征进行脑电分析.近年来单试次脑电分类方法取得了长足进步其中端到端的脑电分类方法能够实现对脑电的自动特征提取和分类但在谎言检测中缺乏研究和应用同时存在无法在测谎场景下直 接应用的问题.本研究设计基于复合反应范式(Complex trial protocol, GTP)进行自我面孔信息识别任务的实验采集了18名被试的脑电数据.研究了不同端到端的单试次ERP分类方法在谎言检测中的应用同时针对单试次脑电解码方法无法直接实际应用的问题提出了一种类自举算法.算法基于数据分布假设通过对比各类刺激图像被视为探针刺激时所训练模型的性能来推断真正的探针刺激.实验结果表明在基于自我面孔信息的GTP的谎言预测中所提出的类自举法性能优于传统探针预测方法，在仅使用少量脑电数据情况下，可实现准确的谎言预测.