Level Up the Deepfake Detection: A Method to Effectively Discriminate Images Generated by GAN Architectures and Diffusion Models

The evolution of generative models has opened new challenges among forensic researchers in creating new sophisticated deepfake detection algorithms. This study analyzed in detail the domain of deepfake detection and recognition by exploring a specialized dataset consisting of real images and synthetic ones generated by nine distinct Generative Adversarial Networks (GANs) and four Diffusion Models (DMs). Then, a multi-level hierarchical strategy was presented to address three different tasks in deepfake detection and recognition: (i) detection between authentic and AI-generated images; (ii) identification of the technology used to create the synthetic data between GANs and DMs; and (iii) recognition of the specific AI-generative architecture. Experimental results showed classification accuracy exceeding 97% in each scenario, outperforming existing methodologies. The obtained models turn out to be robust to various attacks such as JPEG compression and resize, demonstrating that the framework can be used in real-world contexts (e.g. Image Forensics Ballistics).