All too often, deception of one form or another enters into our interactions with others. In certain situations, such as in criminal investigations or matters of national security, it is crucial that deception be detected. Unfortunately, there are no particularly accurate or reliable ways to do so. Polygraphs, often known as “lie detectors,” are notoriously inaccurate and subject to being fooled by a subject that understands the details of their operation. Trained experts, such as law enforcement personnel, do not fare much better. Studies have shown that they are only marginally better at detecting deception than the average person.
A collaboration between researchers at Tel Aviv University and New York University has yielded a device that can detect deceptions, with a high degree of accuracy, using a novel method. The basis for the method is the assumption that deception shows itself through involuntary micro-expressions of the face that are transient and do not match the emotion that the person is trying to convey.
These micro-expressions can be captured by using facial surface electromyography (sEMG), which records the electrical activity of muscles just below the skin surface. Until recently, sEMG systems have been low resolution, cumbersome, unstable, and prone to noise, making them unacceptable for deception detection. However, recent advancements in dry screen-printed electrode arrays have enabled the development of sEMG devices that overcome these shortcomings.
These new sEMG sensors were paired with a machine learning algorithm — a support vector machine classifier. A cohort of forty individuals was selected, and they were asked to pair up. One member of a pair would secretly hear a word via headphones, then they were to either repeat the word, or lie, and tell their partner that another word was spoken. After training the classifier on sEMG data, paired with known classes (truth, lie), the system was evaluated. The average accuracy was found to be a very respectable 73%.
The team found that the individuals in the study showed different types of reactions when telling a lie. Some would show abnormal muscular activity in the cheeks, and for others, it would manifest in the eyebrows. They believe that with a larger sample, a whole host of different give-away indicators would be found. This gives insight into why current techniques tend to perform poorly — they rely on predefined sets of indicators, which presupposes that people share similar indicators of deception. That assumption may be a fatal flaw for those methods.
While the deception detector performed quite well, it was not tested in particularly realistic scenarios. Real deceptions generally include longer statements, consisting of strings of both truth and lies. It is not clear if the method, which was trained on single-word statements, would translate well into use in those scenarios. Further work will be needed in the future to validate this approach in larger, and more realistic, trials.