ANGUS, J. E. and CASTELAZ, P. F. Artificial neural network analysis of polygraph signals. October 1993. Report No. DoDPI93-R-0010. Department of Defense Polygraph Institute, Ft. McClellan, AL 36205.
The purpose of this research was to investigate the use of artificial neural networks (ANN) in classifying psychophysiological detection of deception (PDD) examinations as deceptive or non-deceptive. ANNs are mathematical models of the computing architecture of the human brain. An ANN was designed to accept all four signals (galvanic skin resistance, cardiovascular activity, thoracic respiration and abdominal respiration) from the polygraph output in their entirety. The PDD data used in the study consisted of confirmed Zone comparison Technique (ZCT) examinations of 56 subjects, of which only 15 were non-deceptive. The ANN application resulted in an 87% correct classification of non-deceptive subjects and a 95% correct classification of deceptive subjects. The misclassifications were evenly split: 2 misclassified deceptive (out of 41) and 2 misclassified non-deceptive (out of 15). The two non-deceptive were just slightly over the classification threshold, into the deceptive region of the classification space, and could potentially be called inconclusive. While these results are promising, they are based on a limited set of data, so generalization to a claim that they will successfully address the overall polygraph classification problem requires more extensive evaluation and demonstration on a much larger database of subjects.
Key-words: artificial neural networks, polygraph, signal procession, algorithms, psychophysiological detection of deception
One of the key means of improving the accuracy of the psychophysiological detection of deception (PDD) techniques is computer analysis of PDD data. Computers can analyze factors that are impossible for even the most competent of human examiners to see, no matter how thoroughly he or she inspects the data. Computers can analyze complex waveforms far faster, in much greater detail, and far more consistently than humans.
It is no easy task to determine the best way to analyze the test data. Many statistical approaches have been used, with varying success. The first major approach taken was discriminant analysis to differentiate between innocent and guilty subjects. Other avenues being explored include decision trees, logistic regressions, and fuzzy logic. If we are to find the best approach, we must explore all avenues.
The approach taken in this study is artificial neural network (ANN) analysis. ANNs are a mathematical attempt to mimic the functioning of the human brain, which uses biological neural networks. The conventional computer processes information serially. That is, one operation is conducted after another, sequentially, and each operation is completed before the next is started. On the other hand, the brain processes information in parallel; many operations are going on simultaneously, and the progress of one operation can affect the progress of others. Artificial neural networks also processes in parallel, and are thus able to "learn" how to analyze charts without identifying the criteria for evaluation.
This ability to learn on their own without explicit instructions about what to look for opens the possibility of having computers find novel indices of deception in PDD data. Clearly, this avenue must be investigated if we are to improve the accuracy of PDD decisions.
The procedures used in this study correctly identified 95% of the deceptive subjects and 87% of the truthful subjects. The authors believe this represents the lower bounds on the potential performance of ANNs, as they were limited by a very small amount of data from truthful subjects. The small number of subjects is an important factor limiting the generalizability of the results of this study.