Shallow-FakeFaceNet: A CNN-Based Detection Framework for GAN-Generated and Handcrafted Facial Forgeries

Abstract

With the rapid advancement of digital media technologies, facial image manipulation has become increasingly sophisticated. Both handcrafted editing tools and deep generative models such as Generative Adversarial Networks (GANs) can produce convincingly fake facial images, posing significant threats like misinformation and identity fraud. In this study, we introduce a novel Handcrafted Facial Manipulation (HFM) dataset, containing 1,527 manually edited images across multiple modification types and complexity levels. To detect these fakes along with GAN-generated images, we propose a lightweight neural network called Shallow-FakeFaceNet (SFFN), optimized for low-resolution images (64×64 and 128×128). The detection pipeline includes MTCNN-based face cropping, noise filtering, GAN-based facial super-resolution for enhancing small images, and extensive image augmentation using both Keras and ImgAug. Unlike prior works that rely on fragile metadata, our model operates solely on RGB image data, making it robust against common forgery tactics. Experimental results show that SFFN achieves an AUROC of 72.52% on handcrafted fakes and 93.99% on GAN-generated faces, outperforming several state-of-the-art models. This approach offers a practical, real-world solution for fake media detection in social platforms and biometric verification systems.