Hybrid Deep Learning Approach for Deception Detection on EEG Data Using DWT, FFT and Hyperparameter Tuning

Introduction: Electroencephalography (EEG) is a brain imaging technique that records electrical activity via scalp-attached electrodes, widely used for studying brain functions and diagnosing neurological disorders. Its high temporal resolution makes it ideal for real-time analysis, despite lower spatial accuracy than fMRI or PET. Recent advancements integrating deep learning with EEG have significantly improved applications in areas like lie detection and cognitive research.
Objectives: This study aims to improve the precision of deception detection using EEG data by developing and assessing new deep learning and machine learning algorithms. The research also seeks to compare the performance of these newly proposed methods with existing ones in terms of accuracy and practical use in forensic and investigative applications.
Method: A comparative analysis was conducted to evaluate the accuracy of different machine learning and deep learning techniques for brain fingerprinting and detecting deception based on EEG signals. Models like CNN paired with FFT and DWT were compared to existing algorithms to assess their accuracy improvements.
Findings: The results demonstrated that the newly proposed algorithms, particularly the combination of CNN with FFT and DWT, showed a significant increase in accuracy when compared to current methods. For example, the CNN combined with FFT saw an improvement from 94.12% to 98.89% in accuracy, while the new CNN-DWT-FFT combination reached an impressive accuracy of 99.42%. Although there was a slight drop in accuracy when using CNN with DWT alone, from 98.76% to 98.64%, the proposed models generally outperformed the current algorithms in most scenarios.
Novelty: The unique aspect of this study is the application of deep learning techniques, specifically the combination of CNN with both DWT and FFT, which had not previously been explored for lie detection using EEG data. This innovative approach significantly enhances detection accuracy, making a noteworthy contribution to forensic psychology and law enforcement.

Background & Context

EEG-based deception detection represents a significant evolution from traditional polygraph technology, shifting the focus from peripheral physiological responses to direct measurement of brain activity. EEG's high temporal resolution makes it ideal for real-time analysis, despite lower spatial accuracy than fMRI or PET, and recent advancements integrating deep learning with EEG have significantly improved applications in areas like lie detection and cognitive research.

Brain fingerprinting approaches using P300 event-related potentials have been explored since the late 1980s as a Concealed Information Test variant. However, classification accuracies have varied considerably across studies and methods, with traditional approaches facing challenges in feature extraction and classification robustness. The integration of signal processing transformations like Discrete Wavelet Transform (DWT) and Fast Fourier Transform (FFT) with Convolutional Neural Networks (CNN) represents a novel computational paradigm that may overcome limitations of earlier methods.

This 2025 study by Nagale addresses a critical gap in the literature: while CNNs have been successfully applied to EEG classification tasks, and wavelet/frequency transforms have improved feature extraction in various neurophysiological applications, the systematic combination of CNN with both DWT and FFT specifically for deception detection had not been thoroughly explored prior to this research.

Research Design & Methodology

The study employed a comparative experimental design evaluating multiple machine learning and deep learning architectures for EEG-based deception detection. A comparative analysis was conducted to evaluate the accuracy of different machine learning and deep learning techniques for brain fingerprinting and detecting deception based on EEG signals, with models like CNN paired with FFT and DWT compared to existing algorithms to assess their accuracy improvements.

The methodological approach involved a multi-stage processing pipeline:

• EEG signal acquisition using standard brain fingerprinting/concealed information test protocols
• Signal decomposition using DWT to extract sub-signals across different frequency bands
• Frequency domain transformation using FFT to capture spectral characteristics
• Hybrid CNN architectures designed to process both wavelet-decomposed and frequency-transformed features
• Hyperparameter tuning to optimize model performance across architectures

The study systematically compared three primary configurations: CNN with FFT alone, CNN with DWT alone, and a novel CNN-DWT-FFT hybrid combining both transformation approaches. Performance was evaluated against baseline algorithms using standard classification metrics including accuracy, precision, recall, F1-score, and AUC.

Results & Key Findings

The newly proposed algorithms, particularly the combination of CNN with FFT and DWT, showed a significant increase in accuracy when compared to current methods, with CNN combined with FFT seeing an improvement from 94.12% to 98.89% in accuracy, while the new CNN-DWT-FFT combination reached an impressive accuracy of 99.42%, though there was a slight drop in accuracy when using CNN with DWT alone, from 98.76% to 98.64%.

Key accuracy improvements:

• CNN-FFT: 94.12% → 98.89% accuracy
• CNN-DWT: 98.76% → 98.64% accuracy (slight decrease)
• CNN-DWT-FFT hybrid: 99.42% accuracy (best performing model)

The hybrid CNN-DWT-FFT model achieved the highest classification performance, suggesting that combining time-frequency (DWT) and frequency-domain (FFT) features provides complementary information that enhances the discriminative power of deep learning classifiers for deception detection tasks.

Discussion & Significance

These findings represent a substantial advance in computational approaches to EEG-based deception detection. The 99.42% accuracy achieved by the hybrid model significantly exceeds many previously reported results in the literature, though it's important to contextualize this within controlled experimental conditions rather than field applications.

The success of the CNN-DWT-FFT hybrid likely stems from its ability to capture complementary representations of the EEG signal. DWT provides multi-resolution time-frequency localization particularly suited to non-stationary EEG signals, while FFT captures global frequency content efficiently. The CNN architecture then learns hierarchical features from these dual representations, potentially identifying deception-related patterns that single-transformation approaches miss.

The slight performance decrease observed with CNN-DWT alone (98.76% to 98.64%) suggests that not all transformation-CNN combinations yield improvements, highlighting the importance of systematic empirical evaluation. The study contributes valuable evidence that carefully designed hybrid approaches can outperform single-method architectures, advancing the field's understanding of optimal feature extraction strategies for brain-based deception detection.

Limitations & Considerations

While the study presents impressive accuracy figures, several important limitations warrant consideration. The research appears to have been conducted under controlled laboratory conditions using brain fingerprinting/CIT paradigms, which may not fully represent the complexity and countermeasure vulnerabilities encountered in real-world forensic applications.

Sample size and participant characteristics were not fully detailed in the available information, making it difficult to assess generalizability across diverse populations. The specific EEG dataset utilized, recording parameters, electrode configurations, and stimulus protocols require further specification to enable replication and cross-dataset validation.

High accuracy rates approaching 99-100% in EEG-based detection studies have historically been achieved in optimal laboratory conditions but often decline substantially in field applications. The susceptibility of the proposed CNN-DWT-FFT approach to mental countermeasures, artifact contamination, and individual differences in brain physiology remains to be systematically evaluated.

Practical Applications

If validated across diverse datasets and conditions, the CNN-DWT-FFT hybrid approach could enhance the practical viability of EEG-based concealed information testing for forensic and security applications. The improved accuracy may reduce both false positive and false negative determinations, though real-world implementation would require extensive field testing and validation against known ground truth.

For polygraph examiners and forensic practitioners, this research suggests that neurophysiological approaches integrating advanced signal processing with deep learning may complement or augment traditional psychophysiological detection methods. However, the technology remains in the research domain and faces significant legal, ethical, and scientific hurdles before routine forensic adoption. Practitioners should maintain awareness of these emerging technologies while recognizing that established polygraph techniques remain the current standard for operational credibility assessment.