With the increasing rate of crime, deceit identification has become an issue of concern. The main task here is to classify the EEG data recorded for lie detection, into innocent and guilty. Although various methods have been developed in past to classify EEG data, deep belief nets are rarely used. This paper uses a deep learning technique using the restricted Boltzmann machine with wavelet to obtain the time and frequency domain information of signals. A deep belief network is developed with four RBMs stacked together. Softmax regression is utilized at output layer to classify EEG data into guilty and innocent. An experiment has been performed on EEG data recorded by performing a “Concealed Information Test”. The test consists of a mock crime scenario where some relevant and irrelevant images are presented in front of subjects. The EEG signals generated after flashing these images are recorded and analyzed.