Lie Detection

Lie DetectionVinod Reddy (09005071)Bhanu Prakash (09005050)Hasan Kumar (09005065)by1IntroductionHuman Lie DetectionTechniquesMicro-Expression based designControversy & false-positive results

Outline2Lie detection is the practice of attempting to determine whether someone is lying.Usually this involves asking the subject control questions where the answers are known to the examiner and comparing them to questions where the answers are not known.Introduction3Lie Detectors(though not an accurate name)In event of crime, Can be used in interrogating To find the truthfulness of the evidence.

Why is deception detection important?4May prove useful when hiring potential employees employee theftrevealing whether or not your future spouse/girlfriend truly loves you or is after your moneyDealing with your stock broker, sales rep, lawyer, ex-wife, car dealer, mechanic, scam artist, etc.Why is deception detection important?5To detect whether a person is lying, it is important to know what to look for in the person which shows he is lying.For this we need to know when a lie fails.Lie Detection Techniques 6Two reasons.Failed to adequately prepare a lie. - lack of adequate thinkingInterference of emotions- lack of control on emotionsWHEN DOES A LIE FAIL ? Why lies fails and what behaviour betray a lie.There are

7Two reasons.Failed to adequately prepare a lie. - lack of adequate thinkingInterference of emotions- lack of control on emotionsWHEN DOES A LIE FAIL ?Why lies fails and what behaviour betray a lie.There are

8Lies often fail because of inadequate preparationWhen liar comes up with a lie at the spotMay contradict himselfBeing caught off guard when asked questions which the liar didnt anticipate.Inadequate Preparation9Lies also betrayed by signs of emotionsSimplest case is when the liar fabricate convincingly an emotion which is not felt.Involves concealing his own emotion.Two types of failures 1)some sign of emotion is revealed 2)the liar may produce some inadvertently a deception cue which shows person is lying.

Interference Of Emotions10How do we usually guess whether the other person is saying the truth?Based on the behavior of the person

Eye PatternsCadence of SpeechBody Language of a LiarEmotional Gestures

Lie Detection Techniques ( human )http://www.humanliedetection.com/techniques.php

http://www.apa.org/monitor/julaug04/detecting.aspx11Human lie detection capabilities are limited.For example, a meta-analysis of 253 studies of people distinguishing truths from lies revealed overall accuracy was just 53 percent - not much better than flipping a coin.

Human - conclusionFor example, a recent, as yet unpublished meta-analysis of 253 studies of people distinguishing truths from lies revealed overall accuracy was just 53 percent--not much better than flipping a coin, note the authors, psychologists Charles Bond, PhD, of Texas Christian University, and Bella DePaulo, PhD, of the University of California, Santa Barbara.

12Simple Lie Detector

Build one home easilyhttp://www.aaroncake.net/circuits/lie.asp

a simple lie detector that can be built in a few minutes, but can be incredibly useful when you want to know if someone is really telling you the truth. It is not as sophisticated as the ones the professionals use, but it works. Itworks by measuring skin resistance, which goes down when you lie.

Notes- The electrodes can be alligator clips (although they can be painful), electrode pads (like the type they use in the hospital), or just wires and tape.To use the circuit, attach the electrodes to the back of the subjects hand, about 1 inch apart. Then, adjust the meter for a reading of 0. Ask the questions. You know the subject is lying when the meter changes.

PartTotalQty.DescriptionR1133K 1/4W ResistorR215K PotR311.5K 1/4W ResistorC111uF 16V Electrolytic CapacitorQ112N3565 NPN TransistorM110-1 mA Analog MeterMISC1Case, Wire, Electrodes (See Nots)

13A polygraph is an instrument that simultaneously records changes in physiological processes such as heartbeat, blood pressure, respiration and electrical resistance (galvanic skin response or GSR)The polygraph was invented in 1921 by John Augustus Larson, a medical student at the University of California at Berkeley and a police officer of the Berkeley Police Department in Berkeley, CaliforniaThe underlying theory of the polygraph is that when people lie they also get measurably nervous about lying. The heartbeat increases, blood pressure goes up, breathing rhythms change, perspiration increases, etc.

Polygraph DePaulo and co-author Wendy Morris, a psychology graduate student at the University of Virginia, conducted a meta-analysis into the possible predictors of deception for "Deception Detection in Forensic Contexts" (forthcoming from Cambridge University Press). They warn readers that detecting deception is an inexact science, but note an association between lying and increased pupil size, an indicator of tension and concentration. Second, they find that people listening to liars think they seem more nervous than truth-tellers, perhaps because their voices are pitched higher. And liars are more likely than truth-tellers to press their lips together. On the other hand, they note, liars don't appear to be more fidgety, nor do they blink more or have less-relaxed posture. According to DePaulo and Morris, only when liars are more highly motivated--whenthe stakes are higher--do they seem unusually still and make notably less eye contact with listeners.14

15A baseline for these physiological characteristics is established by asking the subject questions whose answers the investigator knows. Deviation from the baseline for truthfulness is taken as sign of lying.

Polygraph16There are three basic approaches to the polygraph test :-The Control Question Test (CQT): compares physiological response to relevant questions about the crime with the response to questions relating to possible prior misdeedsThe Directed Lie Test (DLT): detect lying by comparing physiological responses when the subject is told to deliberately lie to responses when they tell the truthThe Guilty Knowledge Test (GKT): compares physiological responses to multiple-choice type questions about the crime, one choice of which contains information only the crime investigators and the criminal would know about

Polygraph17Validity : Polygraphy has little credibility among scientists. A 1997 survey of 421 psychologists estimated the test's average accuracy at about 61%, a little better than chance. Critics also argue that even given high estimates of the polygraph's accuracy a significant number of subjects (e.g. 10% given a 90% accuracy) will appear to be lying, and would unfairly suffer the consequences of "failing" the polygraph

PolygraphPolygraph tests- so-called "lie detectors - are typically based on detecting autonomic reactions and are considered unreliable.

That's why psychologists have been cataloging clues to deception - such as facial expressions, body language, writing style and linguistics - to help hook the dishonest.

18Functional magnetic resonance imaging or functional MRI (fMRI) is an MRI procedure that measures brain activity by detecting associated changes in blood flowStudies using fMRI have shown that it has potential to be used as a method of lie detection. While a polygraph detects changes in activity in the peripheral nervous system, fMRI has the potential to catch the lie at the source.

fMRI19The procedure is similar to MRI but uses the change in magnetization between oxygen-rich and oxygen-poor blood as its basic measureThis measure is frequently corrupted by noise from various sources and hence statistical procedures are used to extract the underlying signal

fMRI20The resulting brain activation can be presented graphically by color-coding the strength of activation across the brain or the specific region studiedUsing this method, studies have shown that lies can be distinguished 78% of the time

fMRI21Electroencephalography (EEG) is the recording of electrical activity along the scalp. EEG measures voltage fluctuations resulting from ionic current flows within the neurons of the brainBrain fingerprinting uses EEG to determine if an image is familiar to the subject. This could detect deception indirectly but is not a technique for lie detecting

Brain Observations22Cognitive chronometry, or the measurement of the time taken to perform mental operations, can be used to distinguish lying from truth-tellingBrain-reading uses fMRI and the multiple voxels activated in the brain evoked by a stimulus to determine what the brain has detected

Brain Observations23Truth drugs such as sodium thiopental and marijuana (historically speaking) are used for the purposes of obtaining accurate information from an unwilling subjectInformation obtained by publicly disclosed truth drugs has been shown to be highly unreliable, with subjects apparently freely mixing fact and fantasy

Drugs24Non-invasive lie detection using non-verbal behaviour is performed by the Silent Talker Lie DetectorIt observes and analyses non-verbal behaviour in the form of micro-gestures while a subject is being interviewedIt is grounded in the psychological theory that non-verbal behaviour is modified by a number of influences when a person is being deceptive. These include arousal (in particular stress), cognitive load, duping delight, and behaviour control

Non-Verbal Behaviourhttp://www.apa.org/monitor/julaug04/detecting.aspx

Also investigating bodily deception cues--particularly facial ones--are Ekman and his associates, who in 1978 published the Facial Action Coding System (FACS), which, when combined with voice and speech measures, reaches detection accuracy rates of up to 90 percent, Ekman claims. He and his colleagues are now automating the FACS for use in law enforcement. Meanwhile, they're trying to raise the accuracy rate even higher.Ekman, through close study, learned that "micro-expressions" lasting less than one-fifth of a second may leak emotions someone wants to conceal, such as anger or guilt. At the same time, signs of emotion aren't necessarily signs of guilt. An innocent person may be apprehensive and appear guilty, Ekman points out.He says, "You must use lying as a last interpretation and rule out everything else that's possible."

25At the University of Texas at Austin, psychology professor James Pennebaker, PhD, and his associates have developed computer software, known as Linguistic Inquiry and Word Count (LIWC), that analyzes written content and can, with some accuracy, predict whether someone is lying. Pennebaker says deception appears to carry three primary written markers:Fewer first-person pronouns.Liars avoid statements of ownership, distance themselves from their stories and avoid taking responsibility for their behavior, he says.More negative emotion words,such as hate, worthless and sad. Liars, notes Pennebaker, are generally more anxious and sometimes feel guilty.Fewer exclusionary words,such as except, but or nor--words that indicate that writers distinguish what they did from what they did not do. Liars seem to have a problem with this complexity, and it shows in their writing.

Linguistic Inquiry and Word Count (LIWC)Facial expressions aren't the only clue. Because deception is a social act involving language, researchers are also studying liars' verbal and written output to find distinctive patterns.DePaulo and Morris say that liars take longer to start answering questions than truth-tellers--but when they have time to plan, liars actually start their answers more quickly than truth-tellers. And they talk less. On the whole, to other people, liars seem more negative--more nervous andcomplaining, and less cooperative--than truth-tellers, they say.The content of conversations can be another tip-off. DePaulo andMorris report that liars seem to withhold information, either from guilt or to make it easier to get their stories straight."Liars' answers sound more discrepant and ambivalent, the structure of their stories is less logical, and their stories sound less plausible," they say. Liarsalso use fewer hand movements to illustrate their actions but are more likely to repeat words and phrases, they add.

26Need for it ? Standard polygraph easily fakedNot counted as evidence in courts

Differences from standard polygraphMicro-gesturesAutomatedNo physical contact neededNo trained psycho-physiologist required

Lie Detector Design with fuzzy neural networks(FNN)Need for it? The standard lie detector used today is simply a polygraph to gauge human bodily reactions to questions, the response to which can be faked with some practice.

Early Lie Detectors, (contrary to what their name suggests) do not detect lies but only whether deceptive behavior is being displayed. As a person is questioned about a certain event or incident, the examiner looks to see how the person's heart rate, blood pressure, respiratory rate and electro dermal activity(sweatiness, in this case of the fingers) change in comparison to normal levels. Fluctuations may indicate that person is being deceptive. Polygraph examinations are designed to look for significant involuntary responses going on in a person's body when that person is subjected to stress, such as the stress associated with deception.

What is a micro-gesture?A micro-expression is a brief, involuntary facial expression shown on the face of humans according to emotions experienced. They usually occur in high-stakes situations, where people have something to lose or gain. Unlike regular facial expressions, it is difficult to hide micro-expression reactions. Micro-expressions express the seven universal emotions: disgust, anger, fear, sadness, happiness, surprise, and contempt.Occur within 1/25 and 1/15 of a second. Micro expressions have been found more reliable than polygraph.

27

28Building BlocksAn Artificial Neural Network (ANN)

Camera

ANN:Anartificial neural network(ANN) is an interconnected group of artificial neurons that uses a mathematical or computational model for information processing based on a connectionist approach to computation. In most cases an ANN is an adaptive system that changes its structure based on external or internal information that flows through the network. In more practical terms neural networks are non-linear statistical data modeling tools. They can be used to model complex relationships between inputs and outputs or to find patterns in data.

For systems where there may be data sets of inputs and corresponding outputs, and where the relationship between the input and output may be highly nonlinear or not known at all, we may want to use fuzzy logic to classify the input and the output data sets broadly into different fuzzy classes. Furthermore, for systems that are dynamic in nature the fuzzy membership functions would have to be repeatedly updated. For these types of systems it is advantageous to use a neural network since the network can modify itself.

29

System Flow GraphTODO:

30WorkingPretestFeed data to the systemCase to be investigatedDate and time of the crimeDetails of the crimeSubjects social background, medical history and criminal record (with cooperation of subject)PretestIn the original lie detector, all polygraph testing techniques normally begin with a pre-test interview. The examinee and examiner discuss the test, test procedure, examinees medical history, and details of the test issues. The examiner also observes the behaviour of the examinee and, in test formats that allow for discretion in question design, may gather information to be used in choosing comparison questions for the test. Depending on the complexity of the case, examiner-examinee interactions, and testing technique, the pre-test interview may last for 30 minutes.In this system, the person conducting the examination simply needs to follow a few simple procedures to effectively carry out the same things that the examiner does. These steps are:Enter the case to be investigated.Enter the date and time at which crime has occurred.Enter the details of crime.Enter the subjects social background medical history and criminal record with the cooperation of subject31Data is now scanned and checks for certain vindictive wordse.g. robbed, murder, jailpattern matching techniquesHigher total number of incriminating words found, closer the value of I1 to 1.I1 = fraction of such words found to total such words recognized by system.

WorkingPretest Input VariablePre-test Input variable:The database of criminal and medical records is now scanned by the system. The system checks the database for certain vindictive (Having or showing a strong or unreasoningdesire for revenge) words such as robbed, or murder, or jail, using simple pattern matching techniques. At the end of the scanning, the total number of incriminating (Make (someone) appear guilty of a crime or wrongdoing) words found decides how close to 1 the value of input variable I1 is. Here I1 is the fraction of vindictive words found in the database of the total number of such words recognised by the system. The more the words found, more is the value of I1.

TODO:32Also known as relevant-irrelevant test. Relevant questions real issue of concern to investigatione.g. asking who did it, about evidence, etc.Irrelevant questions provoke no emotionIrrelevant questions are typically asked first.Physiological response of no diagnostic valueGuiltyStronger reaction to relevant questionsInnocentReact similarly

WorkingGeneral testThe general test is also known as the Polygraph and Lie Detection relevant-irrelevant test. As its name implies, the relevant-irrelevant test format compares examinee responses to relevant and irrelevant questions. A relevant question is one that deals with the real issue of concern to the investigation. These questions include asking whether the examinee perpetrate the target act or know who did it and perhaps questions about particular pieces of evidence that would incriminate the guilty person. An irrelevant question is one designed to provoke no emotion.Irrelevant questions are typically placed in the first position of a question list because the physiological responses that follow the presentation of the first question are presumed to have no diagnostic value; they are also placed at other points in the question sequence. Guilty examinees are expected to show stronger reactions to relevant than to irrelevant questions; innocent examinees are expected to react similarly to both question types.

33Convert receiving operator characteristic (ROC) curve/graph (analog signal) to digital signal. Difference of consecutive peaks and lows is taken and averaged out over total number of such differences to give I1i i.e., input variable for the ith response for the general test.WorkingGeneral test Input variablesAfter getting the physiological response from the instrument we need to convert the receiving operator characteristic (ROC) curve/graph which is the analog signal, to digital signal associated with each response. The difference of consecutive peaks and lows is taken and averaged out over the total number of such differences to give I1i. Here I1i is the input variable for the ith response for the general test.

34Comparison question testAsk about general undesirable acts.Peak-of-tension testQuestions are asked in an easily recognized order.A guilty examinee Responsiveness increases as correct alternative approaches in question sequenceDecreases when it has passedOtherse.g. probable-lie and directed-lie comparison tests, known-solution peak-of-tension test

WorkingControl TestComparison question tests (also called control question tests) compare examinees responses to relevant questions to their responses to other questions that are believed to elicit physiological reactions from innocent examinees. Relevant questions are defined as in the relevant-irrelevant test. Comparison questions ask about general undesirable acts, sometimes of the type of an event under investigation. In probable-lie comparison question tests, the instructions are designed to induce innocent people to answer in the negative, even though most are lying. Innocent examinees are expected to experience concern about these answers that shows in their physiological responses. In directed-lie tests, examinees are instructed to respond negatively and untruthfully to comparison questions.The peak-of-tension test is concerned with questions which are asked in an easily recognized order. A guilty examinee is expected to show a pattern of responsiveness that increases as the correct alternative approaches in the question sequence and decreases when it has passed. In a known-solution peak-of-tension test, the examiner knows which alternative is the one truly connected to the incident and evaluates the examinees pattern of responses for evidence of involvement in the incident.

They are such types of questions that basically are answered with "yes", but still are uncomfortable honestly answering. Such an example is "Have you ever lied in order to get out of trouble?" or "Have you ever told a lie?" I'm sure your answer will be "yes". Control questions are those against which your reactions to relevant questions will be compared.

35Convert receiving operator characteristic (ROC) curve/graph to digital signal.Difference of consec. peaks and lows is averaged out to give I2i i.e., input variable for the ith response in the control testIf there are n physiological parameters,Then #input Variables = (2n+1) ( n general test, n control test and 1 pretest)Input variables are fed into neural network (trained beforehand) to generate output.

WorkingControl Test Input VariableAfter getting the physiological response from the polygraph, in this case as well, we need to convert the receiving operator characteristic (ROC) curve/graph which is the analog signal, to digital signal associated with each response. The difference of consecutive peaks and lows is taken and averaged out over the total number of such differences to give I2i. Here I2i is the input variable for the ith response in the control test.If there are n physiological parameters, then our number of input variables becomes 2n+1 (n from the general test, n from the control test, and 1 from pre test input).The greater the values of the input variables, the closer will the final output is to one. However, there is no explicit mathematical relation that defines the final output.These input variables are then fed into the neural network to generate the output. However, as mentioned earlier, the neural network should be initially trained in order to generate a reliable output.

36Fuzzy vs crisp neural networkMembership functions vs weightsObtaining the data set (membership functions of each of the input variables)# data regions = # cases used for training

Generating Membership FunctionIn a crisp neural network, the neural network is used to generate weights. For a fuzzy neural network, however, the neural network is required to generate membership functions.

Obtaining the data set:The data set is nothing but membership functions of each of the input variables. Now how to get these membership functions from crisp values of response variations?Once the sets of data are obtained, the membership values for each of these data points are assigned. While doing so, the type of variation of every response must be known, i.e. whether the response curve is linear, parabolic, exponential, etc. The total possible range of response values must be categorized according to the degree to which that value determines the nature of output. For instance, a particular range of values of GSR may be termed as very high by experts, some values termed as low, others very low and so on. These ranges must be input by the user, from data provided by the experts. Thus, knowing the ranges and the kind of variation of the graph for the response, the membership function for a particular value of the response Ii can be generated. These membership functions become the input data set. The number of data regions = number of cases used for training

TODO: What is membership function, membership values of data points

37Similar to Feed Forward Crisp Neural Network.Sigmoid neuron

TrainingThe next step is to assign quasi random values to different weights connecting the paths between the elements in the layers in the network. Now, we calculate the output using the equation:O=1/(1+exp(-(sigma(XiWi)-t)))The first iteration would be to calculate the output to the layer two.Using the outputs of layer two, we next calculate the output to the layer three.Finally, in the last iteration we calculate the output to layer four, which is the final output of the network in this case.Our next job is to compute the errors for each of the weights and update the weights. While computing the errors we start in the opposite direction as that of computing the outputs.The error is calculated by the formula:Ei = O (calc) O (actual)The errors are distributed to other nodes using the equation:En=On(1-On)sigma(WnjEj)Once the errors associated with every element in the network is known, we can now update the weights.wijk (new) = wijk (old) + a E(i+1)k xjkWhen all the weights of the network have been updated, the input data point is passed through the neural network again and the process is continued until the errors are within acceptable limits. Next the process is repeated with the second data point as well, and so on, till all data points in the training set are used. The performance of the neural network is then checked using the data points in the checking data set. The final weights that we get via the training process are now a constant. These can now be incorporated into the system permanently as the final weights of the system.

38# output in neural network = n+1Todays lie detectors, responses in the form of graph.Fuzzy mathematical expressions must be brought to deal with such situations.(n+1) membership functions combine to give a unique membership function outside neural network, which in turn must be defuzzified to give final outputMin of all membership functions benefit of doubt to the examineeOutputSince the total number of responses is n, thus the number of outputs from the neural network will simply be n+1.However, it must be understood, that in the lie detector used today, the responses are obtained in the form of a graph. These graphs are interpreted by the experts and a combined effect of traces of all the responses is taken into account while judging the validity of the examinees statements.The output cannot be in terms of a clear cut zero or one and there is a certain amount of confusion in the interpretation of these responses. Thus now fuzzy mathematical expressions must be brought in to deal with such situations.Now, when a completely unknown set of data is input, the network generates n+1 membership functions. These membership functions must now be combined to give a unique membership function outside the neural network that must in turn be defuzzified to give the final output.Let the final membership functions be (f1, f2, f3 fn+1). The final membership function must then be the meet of these i.e. f = f1 ^ f2 ^ f3 ^ ^ f(n+1)This implies that the final membership function is the minimum of all these membership functions. In a way this provides benefit of doubt to the examinee.39Mathematically, de-fuzzification of a fuzzy set is a process of rounding it off from its location in the unit hypercube to the nearest vertex.Put simplyFor our system, we propose the value of = 0.5; i.e. for any output greater than 0.5 the output would be 1, otherwise 0.In this case, the case 1 would mean the person is a liar, while 0 would mean the person is truthful.

De-fuzzifying final outputOnce the input for a particular subject is provided, the various input variables are fed into the neural network to give the final output. This output is in the form of a fuzzy set. The final output must however be in crisp form. Mathematically, the defuzzification of a fuzzy set is the process of rounding it off from its location in the unit hypercube to the nearest vertex.For this purpose, we define a lambda cut set matrix.If our final set of outputs be the set A` then the lambda cut set A? , where 0 < = ? < = 1. This set A? is a crisp set and gives the final output. Any element x ? A? belongs to A` with a grade of membership that is greater or equal to the value ?[4]. For our system, we propose the value of ? = 0.5; i.e. for any output greater than 0.5 the output would be 1, otherwise 0. In this case, the case 1 would mean the person is a liar, while 0 would mean the person is truthful.

40Difficult to spot and analyze manually.Requires high processing powers to capture and analyze micro-gestures.

Drawbacks of using micro-expressions41prepare yourself in advance by thinking about what confessions they are looking for, that you can know what things to admit and what things to deny.

Fooling the detectorshttp://sosuave.net/forum/showthread.php?t=166543

http://www.godlikeproductions.com/forum1/message334797/pg1

42What did we discuss?Lie DetectionHuman MethodsTechniques usedSilent Talker DesignControversyThough Lie Detectors are not completely accepted by the scientific community, the day might not be far awayConclusion43Bhattacharjee, Anwesha "An Efficient Lie Detector Using FNN", at the IEEE 7th Student Conference on Research and Development, University Putra Malaysia, 2009. https://sites.google.com/site/fnnliedetector/Charting the behavioural state of a person using a back propagation neural network, Janet Rothwell, Zuhair Bandar, James OShea, David McLean. 2009Simple Lie Detector - http://www.aaroncake.net/circuits/lie.asphttp://sosuave.net/forum/showthread.php?t=166543Use of Fuzzy Set Classification for Pattern Recognition of Polygraph. Knapp, Ulka, jacobs, 1995.

References44Thank youhttp://www.robotshop.com/blog/joke-about-a-lie-detector-robot-583

45