Explaining Neural Network Results by Sensitivity Analysis for Deception Detection

This 2021 study by Zhang and Zhu employed neural networks to detect deception from observers' physiological signals (pupillary responses) rather than suspects' responses. Three architectures (three-layer NN, LSTM, MTL-NN) achieved up to 62% accuracy, surpassing the 54% human baseline. Sensitivity analysis with genetic algorithms successfully extracted interpretable rules from all models, addressing neural network explainability.

Laboratory study training three neural network architectures (standard feedforward, LSTM, and multi-tasking learning networks) on physiological signals from observers viewing video presentations. Rule extraction performed via sensitivity analysis with genetic algorithm-based data reduction.

Background & Context

Human deception detection accuracy typically hovers around 54%, barely better than chance. This persistent limitation has driven researchers to explore whether physiological signals can reveal what conscious judgment misses. The Zhang and Zhu (2021) study represents an innovative approach within deception detection research by focusing not on the person suspected of lying, but on the unconscious physiological responses of observers watching potentially deceptive presentations.

This research builds on prior work demonstrating that observers exhibit measurable physiological reactions when viewing deceptive behavior, even when they cannot consciously identify the deception. The study addresses two critical challenges in applying neural networks to deception detection: achieving accuracy superior to human judgment and making "black box" neural network models interpretable through rule extraction techniques.

Research Design & Methodology

The researchers trained three distinct neural network architectures using physiological signals from observers: a three-layer neural network, a Long Short-Term Memory (LSTM) network, and a Multi-Tasking Learning Neural Network (MTL-NN). This multi-model approach allowed comparison of different architectural strategies for capturing deception-related patterns in physiological data.

The study employed an observer-based paradigm where participants viewed video presentations while their physiological responses were monitored. Key methodological elements included:

• Physiological data collection: Pupillary responses served as the primary physiological signal, capitalizing on the autonomic nervous system's sensitivity to cognitive and emotional states
• Neural network training: Three architectures were trained and evaluated to identify deception from observer responses
• Interpretability enhancement: Sensitivity analysis combined with genetic algorithm (GA) based data reduction was applied to extract human-readable rules from trained models
• Comparative validation: Model performance was benchmarked against baseline human deception detection accuracy

Results & Key Findings

The examined models achieved deception detection accuracy up to 62%, surpassing the average 54% accuracy of human deception detection. This 8 percentage point improvement represents a meaningful advance over unaided human judgment, demonstrating the potential value of physiologically-informed computational approaches.

Key empirical findings included:

• 62% maximum accuracy across the three neural network architectures tested
• Successful rule extraction: Sensitivity analysis combined with genetic algorithm-based data reduction successfully extracted interpretable rules from all three neural network models
• Model generalizability: The rule extraction approach worked across different neural network architectures (standard feedforward, LSTM, and MTL-NN)
• Demonstration that observer physiological responses contain information about presenter veracity that exceeds conscious perceptual judgments

Discussion & Significance

This research makes dual contributions to deception detection science. First, it validates the observer-based physiological approach as a viable alternative to traditional suspect-focused methods. Rather than monitoring the suspected deceiver's responses—which can be subject to countermeasures and individual differences—this approach captures the involuntary reactions of observers who process deceptive cues at an unconscious level.

Second, and perhaps more importantly for practical deployment, the study advances neural network interpretability by demonstrating that sensitivity analysis can extract meaningful rules from trained models. This addresses a critical barrier to adoption: stakeholders in forensic, security, and legal contexts typically require not just predictions, but explanations for those predictions. The combination of sensitivity analysis with genetic algorithm optimization provides a pathway to creating explainable AI systems for deception detection.

The 62% accuracy, while modest in absolute terms, represents a statistically and practically significant improvement over the 54% baseline. In high-stakes contexts such as security screening or investigative interviewing, even moderate improvements in detection accuracy can translate to meaningful operational benefits when combined with other investigative methods.

Limitations & Considerations

The study's reliance on laboratory stimuli—video presentations rather than live interactions—limits ecological validity. Real-world deception occurs in dynamic social contexts with higher stakes, interpersonal dynamics, and contextual cues that cannot be fully replicated in controlled viewing conditions. The 62% accuracy achieved in this constrained environment may not generalize to field applications.

Observer-based approaches introduce practical deployment challenges. Unlike automated analysis of suspect physiological signals, this method requires recruiting, instrumenting, and analyzing multiple observers for each case—a resource-intensive process. The study does not specify how many observers are needed for reliable classification or whether inter-observer variability affects accuracy.

The reliance on pupillary responses, while scientifically sound, constrains application contexts. Pupillometry requires controlled lighting conditions and eye-tracking equipment, making deployment difficult in field settings. Finally, the rule extraction methods, while improving interpretability, may oversimplify the complex patterns learned by neural networks, potentially sacrificing some predictive power for explainability.

Practical Applications

The observer-based approach has potential applications in contexts where traditional polygraph examination is impractical or ethically problematic. For instance, analyzing observer reactions to recorded testimony or public statements could provide a non-invasive screening tool. The interpretable rules extracted via sensitivity analysis could guide human analysts in identifying which physiological patterns most reliably indicate deception, improving both automated and human-assisted decision-making.

For polygraph professionals and researchers, this study underscores the value of exploring alternative physiological indicators and analytical approaches. The genetic algorithm-based rule extraction technique offers a methodology applicable beyond deception detection—any neural network-based diagnostic system could benefit from enhanced interpretability. However, the 62% accuracy ceiling suggests these methods should augment rather than replace traditional investigative techniques and human judgment.