<![CDATA[Purpose – The rapid advancement of Generative AI produces near-perfect synthetic face images, posing fraud risks. This study compares XGBoost and CatBoost for detecting real versus synthetic faces using texture and color features.Methods – Texture features were extracted using GLCM with four orientations (0°, 45°, 90°, 135°), generating 16 features, while YCbCr mean and standard deviation provided 6 color features (22 total). A balanced 750-image dataset (real and AI from GPT Image 2, Nano Banana, Leonardo AI, Canva, Dreamina AI) was split 80:20 and normalized with StandardScaler.Findings – On an asymmetric, source-shifted test split, where real test images were captured independently while synthetic test images came from the same generators as training, CatBoost outperformed XGBoost, achieving 84.67% accuracy (vs. 80.67%), 83.33% precision (vs. 78.05%), 86.67% recall (vs. 85.33%), and 84.97% F1-score (vs. 81.53%), reducing False Positives by 27.78% and False Negatives by 9.09%. However, repeated 5-fold stratified cross-validation (5×10 repeats) revealed only marginal, non-significant differences (0.05-0.87 percentage points, p > 0.09). Research implications – Although the moderate dataset excludes recent architectures such as StyleGAN, findings affirm the feasibility of efficient, interpretable detection through feature engineering. Since within-domain cross-validation shows no significant algorithmic difference, the observed cross-domain gap should be regarded as preliminary, pending validation through repeated cross-domain evaluations across diverse AI platforms.Originality – The novelty lies in comparing XGBoost and CatBoost on multi-orientation GLCM and YCbCr feature fusion for contemporary AI-generated images, identifying cr_std and contrast_0 as stable, interpretable discriminative markers.]]>
Copyrights © 2026
