High-Performance Polygraph-Based Truth Detection System: Leveraging Multi-Modal Data Fusion and Particle Swarm-Optimized Random Forest for Robust Deception Analysis

Deception detection is considered a concern for all individuals in their everyday lives, as it greatly affects human interactions. While multiple automatic lie detection systems exist, their accuracy still needs to be improved, additionally, the lack

Background & Context

The automation of polygraph-based deception detection represents a critical intersection of traditional psychophysiology and modern machine learning. While multiple automatic lie detection systems exist, their accuracy still needs to be improved, and the lack of adequate and realistic datasets hinders the development of reliable systems. This study addresses a fundamental challenge in polygraph science: the need for openly available multimodal datasets and optimized analytical frameworks that can enhance detection accuracy while reducing examiner subjectivity.

The research emerges within a broader trend toward AI-enhanced forensic methods, where machine learning algorithms promise to minimize human bias and standardize polygraph interpretation. Traditional polygraph scoring relies heavily on examiner expertise and subjective judgment, which can be influenced by factors including fatigue, cognitive bias, and inconsistent training. By combining physiological sensor fusion with advanced hyperparameter optimization techniques, this work seeks to establish a reproducible, objective framework for automated deception detection.

Research Design & Methodology

This paper presents a new multimodal dataset with physiological data (heart rate, galvanic skin response, and body temperature), in addition to demographic data (age, weight, and height), collected from 49 unique subjects. The study employed a laboratory-based experimental design where participants were assigned to truthful and deceptive conditions, creating ground truth labels for supervised machine learning.

The methodology integrated multiple components:

• Physiological sensors: Heart rate variability, Galvanic Skin Response (GSR), and body temperature measurements captured during structured questioning protocols
• Machine learning models: Different machine learning algorithms including Random Forest, Logistic Regression, Support Vector Machine, and k-Nearest Neighbor
• Optimization strategies: Grid Search, Bayesian Optimization, and Particle Swarm Optimization (PSO) for hyperparameter tuning
• Validation framework: Robust k-fold cross-validation combined with hyperparameter tuning

The study released the first open multimodal polygraph dataset combining BPM, GSR, body temperature and demographics, along with transparent code, hardware BOM and analysis scripts for full replication. This openness represents a significant methodological contribution, allowing independent verification and extension of the findings.

Results & Key Findings

Random Forest achieved an accuracy of 97%, outperforming Logistic Regression (58%), Support Vector Machine (58% with perfect recall of 1.00), and k-Nearest Neighbor (83%). This dramatic performance difference underscores the importance of algorithm selection for physiological deception detection data.

Key performance metrics included:

• Precision and recall: The model shows excellent precision and recall (0.97 each)
• Processing speed: Computation time of 0.06 s, demonstrating real-time applicability
• Generalization improvement: PSO for hyperparameter tuning substantially reduced the gap between training and validation accuracies from several percentage points to under 1%

The Particle Swarm Optimization approach proved particularly effective at preventing overfitting, a common challenge when working with physiological datasets that exhibit high inter-individual variability. The SVM's perfect recall despite lower overall accuracy suggests it was biased toward predicting the majority class, highlighting the importance of examining multiple performance metrics beyond accuracy alone.

Discussion & Significance

The 97% accuracy figure represents performance on a controlled laboratory dataset and must be interpreted with appropriate caution. Although the Random Forest model attains 97% accuracy on the dataset, the sample size (N = 49) is modest and comparable to other lie detection datasets, and the reported improvement should be interpreted with caution and regarded as a demonstration of potential rather than conclusive evidence of generalisability. This responsible framing acknowledges the limitations inherent in laboratory analog studies.

The primary contribution lies not in claiming definitive superiority over existing methods, but in demonstrating how systematic optimization can enhance performance. The principal contribution is the release of an open multimodal dataset and an optimisation workflow that shows how sensor fusion and hyperparameter tuning can enhance performance on this dataset, with future studies with larger, more diverse samples required to validate these findings.

This work advances the field by providing a transparent, reproducible benchmark. The integration of PSO with Random Forest represents a methodologically sound approach to navigating the complex hyperparameter space of ensemble methods, potentially applicable beyond polygraph contexts to other psychophysiological classification problems.

Limitations & Considerations

The authors acknowledge several critical limitations. The modest sample size of 49 participants limits statistical power and generalizability. Laboratory mock-crime scenarios differ fundamentally from real-world criminal investigations in terms of psychological stakes, stress levels, and consequences of detection—factors known to significantly influence polygraph validity. The demographic composition of the sample and the specific questioning protocols employed are not detailed in available excerpts, making it difficult to assess potential sampling biases.

The absence of field validation represents a major constraint. Laboratory accuracy does not necessarily translate to field performance, as real investigations involve higher emotional arousal, greater motivation to deceive or appear truthful, potential countermeasure use, and examiner-examinee dynamics that cannot be fully replicated in experimental settings. The study does not address countermeasure resistance, a critical concern for operational polygraph applications.

Finally, the reliance on ground truth labels based on experimental condition assignment assumes perfect compliance—participants instructed to lie may not have lied convincingly, and those instructed to tell the truth may have experienced anxiety that mimics deceptive responses. These factors could inflate apparent accuracy.

Practical Applications

The open-source dataset and code provide valuable resources for researchers developing and benchmarking automated scoring algorithms. The demonstrated effectiveness of Particle Swarm Optimization for hyperparameter tuning offers a practical tool that polygraph software developers could integrate into commercial systems. The emphasis on multi-modal sensor fusion reinforces the value of collecting diverse physiological measures rather than relying on single channels.

For polygraph examiners and consumers, this research underscores both promise and caution. While machine learning can potentially reduce subjective bias and improve consistency in scoring, the transition from laboratory to field requires extensive additional validation. Practitioners should view automated systems as decision-support tools requiring expert oversight rather than autonomous diagnostic instruments. The field would benefit from collaborative efforts to create larger, more diverse open datasets that better represent the populations and conditions encountered in operational polygraph screening and criminal investigation contexts.