Distinguishing Between Fake and Real Smiles Using EEG Signals and Deep Learning

Abstract

Genuineness of smiles is one aspect of the field of deception recognition, one that is prevalent in myriad social situations, and it is not easy to tell when a person’s smile is genuine or not for the average person. Machine learning techniques, such as support vector machines or artificial neural networks, can allow better distinction between fake and real smiles by making use of electroencephalograms (EEG) from subjects with a simple experimental protocol, in which the subject’s response is known by the experimenters. Machine learning techniques were previously used in affect recognition, though not for distinguishing real and fake smiles through EEG signals. The objective of this study is to distinguish between fake and real smiles using deep learning techniques, more specifically shallow neural networks, convolutional neural networks, and support vector machines (SVMs) as a baseline from EEG signals. The experimental approach involved presenting subjects with visual stimuli and recording their physical response and their EEG, which was used with the aforementioned algorithms. The SVM classifier used the radial basis function kernel, with optimized parameters, the simple neural network was a three-layer pattern recognition network with 150 hidden units using scaled conjugate gradient as the training function, the convolutional neural networks used stochastic gradient descent with a momentum of 0.95 for all the different architectures, and the optimal one was selected based on the results. The accuracies of the simple neural network, convolutional neural network, and SVM are 88.879 %, 90.446 %, and 48.387 % respectively for subject-dependent classification, and the convolutional neural network yielded 53.418 % for subject-independent classification.