Identification of Reliable Cues for an Automatic Deception Detection System

Abstract

An automatic deception detection system (ADDS) is to detect deceptive human behavior with machine extractable evidences (i.e., cues). One of the most prominent challenges for building a ADDS is the availability of reliable cues. This study represents one of the first attempts to address the system's reliability by identifying the set of reliable cues in order to improve the system performance (detection accuracy).This study addresses two critical challenges of existing machine cues, irreproducibility and inconsistency. First, in order to mitigate the irreproducibility, the study introduces a set of machine measurable cues to estimate the commonality of related machine cues. These more reproducible cues are referred to as the macro cues which can be applied for automatic pattern recognition. Second, in order to address the consistency, the study separates cues based on the controllability, and defines the strategic cues as those can easily be manipulated by deceivers during interaction. The strategic cues fluctuate during deception and thus are less consistently reliable as predictors for the ADDS. On the contrary, the nonstrategic cues are more consistent. This study also considers other moderator effects that influencing the ADDS performance: time and the condition of interviewer's immediacy (ERIMD).The macro cues are automatically estimated from the micro cues based on the predefined relational models. The empirical data support the relationship models between macro and micro cues. Results show that macro cues mitigate the irreproducibility problem by reducing the variability in the single cues. However, the results also show that using macro cues as predictors in the discriminant analysis does not perform better than micro cues, and thus imply the needs to adjust weights of important components when constructing the macro cues. In terms of the consistent cues, results show that the nonstrategic cues are relatively more consistent than strategic ones in ADDS performance. Furthermore, the study suggests that particular detection methods must be tailored according to the feature of strategic and nonstrategic cues. The findings have many potential implications. One is to use the macro cues to recognize the dynamic patterns in deceptive behaviors. Specifically, truthtellers increase the certainty, immediacy, and tend to decrease the cognitive load; but deceivers behave the opposite. The other is to rely on the characteristics of strategic cues to manipulate the communication environment to improve the ADDS performance. This concept is also referred to as the Proactive Deception Detection (PDD). In the current study, the interviewer's immediacy is a controllable environment factor for PDD. The high ERIMD increase the system performance because it has higher overhead added to the deceptive behavior to trigger more abnormal cues. In sum, methods and results of this study have multiple impacts in information assurance and human-computer interaction.