Chinese Deception Detection Based On Deep Belief Network

Deception detection is becoming indispensable to a growing number of applications such as preventing phone scams, criminal detection and identification potential deception from informants at intelligence department etc. There are various ways for deception detection, including pulse and blood pressure based methods, facial expression based methods, brain wave based methods and speech signal based methods. Among all these methods, speech signal based methods are those which leverage speech signal process techniques to predict whether people tell lies. Due to its convenience and non-contact, speech signal based methods gain more attention and have certain advantages compared with other approaches.However, research on speech based deception detection is still in the beginning stage and poses unique technical challenges. Firstly, unlike English deception detection, there are no public speech corpora in Chinese. Lack of real-world useful Chinese deception detection corpus cramps the development of techniques in deception detection research. Secondly, existing methods mostly use time domain or frequency domain features to train supervised models. However, manufacturing feature engineering is time-consuming and not enough to capture the complex phenomena in speech signals. Faced with these problems, the following solutions are proposed:1. To overcome the lack of corpus, firstly, some corpus construction principles that can guide the construction process are designed. Then subjects are motivated to tell lies with money rewards as driven force. This makes our corpus quite rich and varied. Moreover, a perception experiment is applied to prove the validity of our corpus.2. A deep learning approach to learn the speech representation using basic features through Deep belief Network is proposed. Most existing approaches only focus on single speech feature or the combination of multi features. The representation ability of these basic features is limited. Our model takes the basic features as input and the learnt deep features can capture the complex information in speech signals. Experimental results show that our learnt deep features can outperform basic features when combined with some existing classifiers.3. A cross-gender deception detection method based on TrAdaBoost is proposed. Since traditional machine learning algorithms cannot be directly applied to perform a cross-gender deception detection due to the distribution diversity on different genders, this paper explores the problem of cross-gender deception detection and proposes a transfer-learning based approach to solve this problem. Compared with the baseline methods, this approach can significantly improve the performance of cross-gender deception detection.