Abstract: There is great significance in lie detection for the criminal investigation and psychological disease treatment. To distinguish lying, thirty subjects were divided into lying and telling-truth groups randomly and three groups of nonlinear features——complexity measures including Kolmogorov complexity, approximate entropy and sample entropy were extracted. By statistical analysis, the feature vector was constructed by using complexity on the muilti electrodes with significant difference of complexity values between the two groups of subjects. The support vector machine was used to classify and idendify feature samples. The study finds that there are more electrodes with significant difference of complexity values for the sample entropy, and the highest classification accuracy can be observed for the feature vector constructed from the sample entropy, compared with the other two featues. Experimental resutls indicate that sample entropy could be used to classify EEG signals in lying from EEG signals in telling-truth, which provides a new alternative for EEG-based lie detection method.