A Review of Automatic Lie Detection from Facial Features

The growth of machine learning and artificial intelligence has made it possible for automatic lie detection systems to emerge. These can be based on a variety of cues, such as facial features. However, there is a lack of knowledge about both the development and the accuracy of such systems. To address this lack, we conducted a review of studies that have investigated automatic lie detection systems by using facial features. Our analysis of twenty-eight eligible studies focused on four main categories: dataset features, facial features used, classifier features and publication features. Overall, the findings showed that automatic lie detection systems rely on diverse technologies, facial features, and measurements. They are mainly based on factual lies, regardless of the stakes involved. On average, these automatic systems were based on a dataset of 52 individuals and achieved an average accuracy ranging from 61.87% to 72.93% in distinguishing between truth-tellers and liars, depending on the types of classifiers used. However, although the leakage hypothesis was the most used explanatory framework, many studies did not provide sufficient theoretical justification for the choice of facial features and their measurements. Bridging the gap between psychology and the computational-engineering field should help to combine theoretical frameworks with technical advancements in this area.