Inspired by the work of psychologists who study the human face
for clues that someone is telling a high-stakes lie, University at Buffalo (UB)
computer scientists are exploring whether machines can also read the visual
cues that give away deceit.
Results so far are promising: In a study of 40 videotaped
conversations, an automated system that analyzed eye movements correctly
identified whether interview subjects were lying or telling the truth 82.5% of
the time.
That’s a better accuracy rate than expert human interrogators
typically achieve in lie-detection judgment experiments, said Ifeoma Nwogu, a
research assistant professor at UB’s Center for Unified Biometrics and Sensors
(CUBS) who helped develop the system. In published results, even experienced
interrogators average closer to 65%, Nwogu said.
“What we wanted to understand was whether there are signal
changes emitted by people when they are lying, and can machines detect them?
The answer was yes, and yes,” said Nwogu.
The research was peer-reviewed, published, and presented as part
of the 2011 IEEE Conference on Automatic Face and Gesture Recognition.
Nwogu’s colleagues on the study included CUBS scientists Nisha
Bhaskaran and Venu Govindaraju, and UB communication professor Mark G. Frank, a
behavioral scientist whose primary area of research has been facial expressions
and deception.
In the past, Frank’s attempts to automate deceit detection have
used systems that analyze changes in body heat or examine a slew of involuntary
facial expressions.
The automated UB system tracked a different trait—eye movement.
The system employed a statistical technique to model how people moved their
eyes in two distinct situations: during regular conversation, and while
fielding a question designed to prompt a lie.
People whose pattern of eye movements changed between the first
and second scenario were assumed to be lying, while those who maintained
consistent eye movement were assumed to be telling the truth. In other words,
when the critical question was asked, a strong deviation from normal eye
movement patterns suggested a lie.
Previous experiments in which human judges coded facial
movements found documentable differences in eye contact at times when subjects
told a high-stakes lie.
What Nwogu and fellow computer scientists did was create an
automated system that could verify and improve upon information used by human
coders to successfully classify liars and truth tellers. The next step will be
to expand the number of subjects studied and develop automated systems that
analyze body language in addition to eye contact.
Nwogu said that while the sample size was small, the findings
are exciting.
They suggest that computers may be able to learn enough about a
person’s behavior in a short time to assist with a task that challenges even
experienced interrogators. The videos used in the study showed people with
various skin colors, head poses, lighting and obstructions such as glasses.
This does not mean machines are ready to replace human
questioners, however—only that computers can be a helpful tool in identifying
liars, Nwogu said.
She noted that the technology is not foolproof: A very small
percentage of subjects studied were excellent liars, maintaining their usual
eye movement patterns as they lied. Also, the nature of an interrogation and
interrogators’ expertise can influence the effectiveness of the lie-detection
method.
The videos used in the study were culled from a set of 132 that
Frank recorded during a previous experiment.
In Frank’s original study, 132 interview subjects were given the
option to “steal” a check made out to a political party or cause they
strongly opposed.
Subjects who took the check but lied about it successfully to a
retired law enforcement interrogator received rewards for themselves and a
group they supported; Subjects caught lying incurred a penalty: they and their
group received no money, but the group they despised did. Subjects who did not
steal the check faced similar punishment if judged lying, but received a
smaller sum for being judged truthful.
The interrogators opened each interview by posing basic,
everyday questions. Following this mundane conversation, the interrogators
asked about the check. At this critical point, the monetary rewards and
penalties increased the stakes of lying, creating an incentive to deceive and
do it well.
In their study on automated deceit detection, Nwogu and her
colleagues selected 40 videotaped interrogations.
They used the mundane beginning of each to establish what
normal, baseline eye movement looked like for each subject, focusing on the
rate of blinking and the frequency with which people shifted their direction of
gaze.
The scientists then used their automated system to compare each
subject’s baseline eye movements with eye movements during the critical section
of each interrogation—the point at which interrogators stopped asking everyday
questions and began inquiring about the check.
If the machine detected unusual variations from baseline eye
movements at this time, the researchers predicted the subject was lying.