<p>Electroencephalography (EEG)-based deception detection provides a promising alternative to conventional polygraph methods by directly measuring neural activity with millisecond-level temporal precision. This paper presents an end-to-end framework that integrates a rigorous preprocessing pipeline with a graph convolutional neural network (GCNN)-based feature extractor and a deep learning classifier to distinguish truthful from deceptive responses. EEG signals were segmented into overlapping epochs to preserve temporal dynamics and increase the number of training samples. To prevent data leakage, all epochs originating from the same trial were assigned exclusively to a single data partition. A two-stage normalization strategy was employed, consisting of subject-wise standardization to reduce individual baseline differences and domain-invariant normalization to minimize inter-subject variability. To improve robustness and generalization, Gaussian noise injection and channel dropout were incorporated as data augmentation techniques. Unlike conventional convolutional neural networks (CNNs), which primarily capture local temporal dependencies, the proposed GCNN explicitly models the EEG electrode configuration as a graph, where nodes represent channels and edges encode spatial proximity and functional connectivity. Through graph convolution operations, information is propagated across neighboring electrodes, enabling the extraction of both local and global neural patterns relevant to deception detection. The classification stage employs a multi-layer perceptron (MLP) that transforms the learned GCNN representations into class probabilities, with dropout regularization and early stopping applied to mitigate overfitting. To evaluate cross-subject generalization, a leave-one-subject-out (LOSO) validation strategy was adopted, where data from each subject were held out for testing while the model was trained on the remaining subjects. Performance was assessed using accuracy, precision, recall, and F1-score across all LOSO folds. Experimental results demonstrate that the proposed framework effectively captures discriminative spatiotemporal EEG patterns with 81% accuracy on a private dataset and 84.9% accuracy on the publicly available LieWaves dataset, and provides robust performance for subject-independent deception detection, highlighting the potential of GCNN-based architectures for EEG-based lie detection.</p>