Hybrid CNN-based detection of forged facial images using Gabor filters and local binary patterns

Abstract

This study proposes a hybrid deep learning framework for detecting face morphing and swapping attacks using texture descriptors and advanced CNN classifiers. The process begins with the Viola-Jones Haar cascade, a widely used face detection method that effectively identifies and isolates facial regions. Subsequently, Local Binary Patterns (LBP) and Gabor filters are applied to extract texture-based features from these facial regions. Finally, the classification stage employs Simple Convolutional Neural Networks (CNN), Custom CNN and ResNet-152 to effectively distinguish between authentic and tampered images. We integrate LBP and Gabor filters to extract robust texture features and classify them using a combination of Simple CNN, Custom CNN, and ResNet-152. Our approach achieves an ACER of 2.82% on the DEFACTO dataset, outperforming several state-of-the-art methods. Additionally, the paper presents ablation studies that confirm the complementary benefits of LBP and Gabor features in enhancing detection performance and benchmarking against five state-of-the-art methods, followed by a detailed analysis of model performance.