Hierarchical Multi-Class Deepfake Detection Using EfficientNetB4 for Source Attribution

Department of Computer Science, Central University of Rajasthan, Ajmer - 305817, Rajasthan, India

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

The rising of deepfake media (images, videos and audio etc.) has become a major issue in the digital era. Deepfake generation techniques such as Generative Adversarial Networks and Diffusion Models, shows major challenge to individual’s privacy, democracy and national security. While the previously deepfake detection techniques deals with binary classification such as real and fake, it remains explored to detect deepfakes with multiple generation techniques concurrently. Despite the recent improvements, no existing work has addressed the problem of mapping all three major generation families (GAN-based, diffusion-based, and face manipulation techniques) into a single hierarchical pipeline, which is a crucial research gap. In order to overcome these limitations this paper we present a multi- class deepfake detection framework capable of distinguishing between real images and nine distinct synthesis methods: DALL-E, Face2Face, FaceSwap, StyleGAN, NeuralTextures, Stable Diffusion, DeepFaceLab, FaceShifter, Midjourney. We used transfer learning with EfficientNetB4 as backbone and proposed hierarchy of three levels and achieves 98.89 % and 98.52% accuracy for binary and multiclass attribution, respectively. The framework further achieves 99.31% precision for binary classification and a macro-average AUC of 0.9995 for multiclass attribution. Our confusion matrix analysis shows that the model is able to differentiate between the various generation strategies and that it is particularly successful in finding diffusion-based approaches and traditional face manipulation techniques. The hierarchical structure offers actionable attribution for forensic use, enabling applications in content moderation, forensic investigations, and media verification.