Bi-GRU Based Deception Detection using EEG

Background & Context

Deception detection represents a critical challenge across multiple domains including security screening, criminal investigations, and forensic psychology. The Bag-of-Lies dataset captures spontaneous, self-initiated deception in realistic scenarios, unlike many prior datasets relying on forced or artificial deception, filling a significant gap in naturalistic deception research.

Traditional polygraph techniques rely on autonomic physiological responses, but EEG data is employed to investigate the cognitive aspects related to deception. While EEG-based deception detection has shown promise, studies on deception detection using EEG data are scarce in current literature, making this contribution particularly timely.

This 2025 study by Avola and colleagues advances the field by applying bidirectional temporal modeling to brain activity patterns during deception, leveraging deep learning architectures specifically designed to capture the sequential nature of neural signals associated with truthful versus deceptive responses.

Research Design & Methodology

The study used the Bag-of-Lies dataset, a multimodal dataset tailored for deception detection that includes video, audio, gaze, and EEG data registered from 35 unique subjects. Set A of the dataset contains data from 22 unique users for whom all four data modalities are available, consisting of 201 recordings, 108 of which are truths and 93 are lies.

EEG data with 13 channels was collected using an EPOC+ headset. The preprocessing pipeline included several key steps:

• Bandpass filtering to remove noise and isolate relevant frequency components
• Segmentation using overlapping windows of 64 time steps
• Class balancing via undersampling to address the imbalance between truth and lie samples
• Gaussian noise augmentation with a noise factor of 0.02 to generate additional synthetic samples
• 80% training and 20% testing split

The neural architecture employed a Bi-GRU model with three bidirectional layers processing information in both forward and backward directions. The three stacked Bi-GRU layers with dropout regularization produced hidden representations of size (64,256), (64,128), and (64,128), followed by fully connected dense layers with dropout for regularization. Each layer was followed by dropout at a 0.5 rate to prevent overfitting.

Results & Key Findings

The model achieved a test accuracy of 97% with a test loss of 0.1696, representing a substantial improvement over prior methods on the same dataset.

Detailed performance metrics demonstrated strong discrimination between classes:

• Precision: 97% for both lies and truths
• Recall: 96% for lies and 97% for truths
• F1-score: 97%

The study's performance significantly exceeded previous approaches on the same dataset. The original dataset paper achieved 58.71% accuracy using Random Forest, while LieNet achieved 95.91% accuracy by converting EEG data into images and using a CNN approach. A previous Bi-LSTM analysis on the same dataset provided the highest single-modality accuracy of 82.24%.

Discussion & Significance

The study achieved 97% accuracy, outperforming several state-of-the-art models applied to the same dataset, including Random Forest, Bi-LSTM, and CNN-based approaches. This advancement demonstrates that bidirectional temporal modeling can effectively capture the sequential cognitive processes underlying deception.

The dataset stands out in existing literature as it closely mirrors real-world scenarios of deception without relying on Event-Related Potentials (ERP), allowing participants the freedom to fabricate a deceitful narrative. This naturalistic approach enhances ecological validity compared to traditional laboratory paradigms where deception is artificially induced.

The success can be attributed to data augmentation techniques, preprocessing of EEG signals, and the application of advanced deep learning architectures such as Bi-GRU. The bidirectional architecture's ability to process temporal information in both directions appears particularly well-suited to detecting the complex neural signatures associated with cognitive load and inhibitory control during deception.

Limitations & Considerations

Some errors were observed in the form of false positives and false negatives, which could be linked to ambiguous signals that may not fit neatly into either class, and the limited size of the dataset could contribute to limiting the generalization capabilities of the model.

The study used only 22 participants with complete EEG data, yielding 201 total recordings. This relatively small sample size raises questions about generalizability across diverse populations and deception contexts. The naturalistic, self-selected deception scenario, while ecologically valid, may differ from high-stakes real-world deception where motivational factors and cognitive load vary substantially.

Additionally, the study focused exclusively on the EEG modality. While this isolates cognitive processes, real-world applications may benefit from multimodal integration combining brain activity with behavioral and physiological channels.

Practical Applications

The results demonstrate the effectiveness of bidirectional temporal modeling for EEG-based deception detection and suggest potential for real-time applications. The relatively lightweight nature of consumer-grade EEG headsets compared to traditional polygraph equipment could enable more accessible screening tools.

For practitioners, this research demonstrates that modern deep learning can extract meaningful deception-related patterns from portable EEG devices. However, the 97% accuracy, while impressive in research settings, still produces a 3% error rate that would require careful consideration in high-stakes forensic or security contexts. The technology shows promise for augmenting human judgment in investigative interviews, pre-employment screening in permissible contexts, and research applications, though ethical and legal frameworks would need development before operational deployment.