Escaping Generator Bias: A One-Class Classification Framework for Generalisable AI-Generated Face Detection

Abstract

In recent years, previous studies indicate a shift from Generative Adversarial Networks (GANs) to diffusion-based models, which exposed some limitations to traditional AI-generated face or image detectors. Some of these detectors are known to struggle in generalising beyond generators they were trained on. This study takes a one-class classification (OCC) approach that models the distribution of real facial images to truly understand what makes a face real or not. Using a frozen CLIP visual encoder combined with a hybrid Deep SVDD and reconstruction-based scoring head, we trained our model exclusively on real faces from the FFHQ dataset and compared it against a supervised binary classifier. It was further evaluated under a strict 5% False Acceptance Rate across three unseen diffusion generators. The OCC framework achieved notably higher recall than the binary baseline, including a 40-point improvement on Midjourney, while the baseline retained an edge on generators resembling its training data. These findings support one-class modeling as a promising direction for building more generalisable deepfake detectors. Future studies should explore more diverse authentic training datasets and more flexible normality models to further enhance detection capabilities. This study offers valuable insights into the intersection of AI and authenticity verification, contributing to efforts at maintaining the integrity of digital media in an era of increasingly sophisticated AI-generated content.