A Review on Robustness of Deception Detection System Based on Eye Movement Features

Deception detection has become an essential field with applications in law enforcement, security, and business. Among various approaches, eye movement features (EMF) have gained increasing attention because of their ability to reveal subtle psychophysiological cues associated with deception. However, the use of EMF to detect deception is still hampered by varying reliability in its application and a lack of analysis of the robustness of these features. To address this research gap, this study presents a systematic literature review (SLR) of studies published between 2019 and August 2025 that investigated deception detection systems (DDS) utilizing EMF. A total of 29 publications were analyzed using robustness parameters. The findings show that most studies achieved only low to medium robustness, with multimodal approaches demonstrating higher robustness score than unimodal approaches. Major challenges include limited dataset diversity, insufficient subject variation, and high computational requirements for advanced validation methods such as Leave-One-Out Cross Validation (LOOCV). This study highlights the potential of combining multimodal approaches with deep learning (DL) and transfer learning techniques, supported by data augmentation and modality weighting. These strategies are expected to improve the generalization, accuracy, and fairness of DDS based on EMF.

Systematic literature review of 29 publications from 2019 to August 2025 examining deception detection systems utilizing eye movement features, evaluated using robustness parameters including dataset diversity, subject variation, validation methodology, and computational requirements. Studies were categorized as unimodal versus multimodal approaches.

Background & Context

Deception detection represents a critical challenge across law enforcement, security screening, and forensic contexts. While traditional polygraph methods rely on autonomic physiological responses, eye movement features (EMF) have emerged as a promising complementary modality for detecting deception due to their sensitivity to cognitive load, attention allocation, and information processing differences between truthful and deceptive responses.

Despite growing interest in eye-tracking-based deception detection systems (DDS), the field has lacked systematic analysis of the robustness and generalizability of these approaches. This systematic literature review addresses a critical gap by evaluating not just the accuracy of EMF-based deception detection, but the methodological rigor and reliability that determine whether these systems can transition from laboratory proof-of-concept to real-world application. The review's focus on robustness parameters represents an important evolution beyond simple accuracy reporting toward assessment of system resilience across varied conditions and populations.

Research Design & Methodology

The authors conducted a systematic literature review (SLR) following established protocols to identify and analyze studies published between 2019 and August 2025 that investigated deception detection systems utilizing eye movement features. The six-year timeframe captures the recent surge in machine learning and deep learning applications to psychophysiological deception detection.

The review ultimately analyzed 29 publications using robustness parameters as the primary evaluation framework. Key robustness dimensions likely included dataset diversity, participant variability, cross-validation methodology, generalization across conditions, and computational efficiency. Studies were categorized as unimodal (using only eye movement features) versus multimodal (combining EMF with other physiological or behavioral modalities).

The analysis specifically examined several critical validation approaches, including Leave-One-Out Cross-Validation (LOOCV), which tests generalization to unseen individuals—a crucial requirement for real-world deception detection applications where systems must work on new examinees without retraining.

Results & Key Findings

The systematic review revealed that most EMF-based deception detection systems achieved only low to medium robustness scores, indicating significant limitations in their reliability and generalizability despite often-impressive accuracy figures reported in individual studies.

Critical findings included:

• Multimodal approaches demonstrated higher robustness scores than unimodal approaches, suggesting that combining eye movement data with other modalities (such as physiological signals, facial expressions, or voice features) produces more resilient deception detection systems
• Limited dataset diversity emerged as a major weakness, with many studies relying on small, homogeneous samples that limit generalizability across populations, cultures, and contexts
• Insufficient subject variation constrained the ability of systems to perform reliably on individuals with different demographic characteristics, cognitive abilities, or baseline eye movement patterns
• High computational requirements for advanced validation methods like LOOCV presented practical barriers to rigorous testing, particularly for deep learning approaches requiring extensive training time

Discussion & Significance

This systematic review provides a sobering assessment of the current state of eye movement-based deception detection. While individual studies frequently report promising accuracy rates, the low-to-medium robustness scores reveal that these systems often fail to demonstrate the reliability necessary for operational deployment. The finding that multimodal approaches outperform unimodal systems aligns with broader trends in deception detection research showing that no single indicator reliably discriminates truth from deception across all contexts.

The authors' identification of specific technical pathways forward—combining multimodal approaches with deep learning and transfer learning techniques, supported by data augmentation and modality weighting—offers concrete directions for advancing the field. Transfer learning is particularly promising for addressing limited dataset diversity, as models pretrained on large datasets could be fine-tuned on deception-specific data. Modality weighting could dynamically adjust the contribution of different features based on their reliability in specific contexts or for specific individuals, potentially improving both accuracy and fairness.

Limitations & Considerations

As a systematic review, this study inherits the limitations of its included primary studies. The predominance of laboratory-based research with mock crime scenarios limits ecological validity—real-world deception involves higher stakes, different motivational structures, and more complex social dynamics than laboratory paradigms typically capture. The review period of 2019-2025 may have excluded earlier foundational work that established important methodological principles, though it effectively captures the recent machine learning revolution in this domain.

The finding of insufficient dataset diversity and subject variation reflects a field-wide challenge in psychophysiological research: collecting large, diverse datasets with ground truth labels for deception is ethically and practically complex. Real-world deception is difficult to verify with certainty, and laboratory deception may not generalize to field contexts. Additionally, the computational demands of rigorous validation noted in the review may inadvertently create a publication bias toward simpler models tested with less demanding validation approaches.

Practical Applications

For practitioners and policymakers, this review's findings counsel caution in deploying EMF-based deception detection systems operationally. The low-to-medium robustness scores suggest current systems may perform unreliably when confronted with populations, contexts, or individual differences not well-represented in training data. This has significant implications for fairness and accuracy in security screening, criminal investigation, and other high-stakes applications where false positives and false negatives carry serious consequences.

The review points toward multimodal integration as the most promising near-term path forward. Examiners and system developers should prioritize approaches that combine eye movement features with established polygraph measures, voice analysis, or behavioral indicators rather than relying on eye tracking alone. The emphasis on data augmentation and transfer learning techniques suggests that investment in larger, more diverse datasets and in adaptation methods that improve generalization should be research and development priorities for organizations seeking to develop or deploy these technologies responsibly.