Deception detection, a single case study

Deception detection, a single case study

Video footage of Lance Armstrong denying the use of doping provides unique material for a single case study of deception.

Using social signal processing techniques (i.e., expression recognition software in combination with machine learning) we analysed a collection of video sequences of Lance Armstrong in an attempt to determine visual cues of deception. A preliminary analysis performed on a small collection in January 2013 suggested that we could predict deception with about 80% accuracy. The preliminary study was prompted by Nik Wouters, a journalist of NOS Journaal (the main Dutch news programme) and resulted in a small item (in Dutch).

Initial experiment

The 80% claim was a rough estimate based on a rather small sample of video sequences of Armstrong denying his doping use (deception instances) and of Armstrong talking about his fight against cancer (no deception instances). The item gave rise to a constructive discussion with Pepijn van Erp of Stichting Skepsis a Dutch organisation dedicated to the practice of scientific skepticism. Van Erp raised questions regarding the validity of the 80% claim.

Full experiment

In 2013, we replicated and extended the preliminary analysis on an extended collection of video fragments. Our master student Steven van Leer collected the data and performed the initial analyses. In total the collection consisted of 62 thin slices (short excerpts) from video interviews with Armstrong. Fifteen slices were taken directly after the interviewer asked Armstrong about his doping use. These were labelled as deceptive slices. The remaining 47 slices were selected after questions about his illness and were labelled non-deceptive. Using a digital tool for the automatic analysis of the facial action units (the building blocks of facial expressions), we measured the activity of parts of the face for each of the slices. Our results corroborate our earlier estimates. The prediction performance turned out to vary between 70 and 80%, depending on the action units employed. In subsequent analysis we narrowed down the most predictive action unit to be action unit 4, the so-called brow lowerer.  This single action unit predicted deception with an accuracy of 83%.  In the case of Armstrong, the appearance of vertical wrinkles on his forehead (one of the cues for action unit 4) is quite prominent.

Interpretation of results

What do our results imply? Are we able to detect deception? No, certainly not. The proper interpretation of our results is that we are able to predict with an accuracy of 70-80% whether or not a short fragment of Armstrong in a TV interview is associated with a deceptive statement. Visual reassessment of the deceptive fragments gives the impression that, whenever confronted with doping acquisitions, Armstrong developed an (what seems to be) acted response. It may very well be the case that this acted response is highly context dependent. In other words, in another setting deception may express itself in a totally different way. More importantly, there is strong reason to believe that deceptive behaviour is highly idiosyncratic (see, e.g., Porter and ten Brinke, 2010).

To summarise, our findings apply to a single person (Armstrong) in a quite specific setting (TV interviews), and cannot be readily generalised to other persons or situations. Despite this limitation, our single-case study reveals one of the many styles of deception thanks to the relatively uniques availability of frontal-face videos of a deceiving person.

 

 

Porter, S. and ten Brinke, L. (2010), The truth about lies: What works in detecting high-stakes deception?. Legal and Criminological Psychology, 15: 57–75. doi: 10.1348/135532509X433151