How Well Do Simple Features Detect Fake Faces? A Comparison with Deep Learning

Abstract

The recent proliferation of highly realistic synthetic facial images, enabled by generative adversarial networks (GANs) and diffusion based models, presents a growing challenge to digital media authenticity and visual forensics. This study investigates the problem of distinguishing real from AI-generated faces through a comparative evaluation of both traditional machine learning techniques and modern deep learning architectures. A custom dataset was constructed, combining real facial images from the FFHQ dataset with synthetic counterparts generated via the “This Person Does Not Exist” platform. We extracted local binary pattern (LBP) descriptors and trained multiple classifiers (including Random Forest,K-Nearest Neighbor, Support Vector Machine, Logistic Regression, and Naive Bayes) providing a robust baseline for interpretable face authenticity detection. In parallel, we evaluated convolutional neural networks and EfficientNet models, both with and without fine-tuning. Our results demonstrate that lightweight LBP-based approaches can yield competitive accuracy, offering interpretable and efficient alternatives to deep models in constrained settings.We further discuss the implications of these findings in light of current advances in generative models, and propose directions for enhancing detection robustness across unseen generative methods.