Artificial intelligence and deepfake technologies have enabled the rapid generation of synthetic text, images, audio, and video, creating significant challenges in detecting fraudulent content. In healthcare, fake medical data can compromise patient safety, cause misdiagnoses, and erode trust in clinical systems. Existing literature shows that relatively few studies have addressed medical deepfake detection as a multiclass classification problem, particularly in distinguishing between real images and multiple categories of synthetic images generated by both diffusion-based and GAN-based models. This study investigates deep learning–based approaches for multiclass medical deepfake image detection using publicly available skin cancer datasets. The problem is formulated at two levels: (i) generator-level classification, where real images are distinguished from specific generators, and (ii) family-level classification, where images are categorized as Real, GAN-based, or Diffusion-based. To construct the multiclass setting, synthetic images were generated using diffusion-based methods, including Denoising Diffusion Probabilistic Models (DDPM) and Vector Quantised Diffusion (VQ-Diffusion), as well as GAN-based approaches such as Pix2Pix GAN and DCGAN. The study evaluates multiple deep learning architectures, including InceptionV3, MobileNetV2, ConvNet, and AlexNet, for multiclass deepfake detection. Furthermore, a hybrid transfer learning framework combining complementary feature representations was proposed to enhance prediction performance. To improve adaptability across generator families and reduce distribution shifts between real and synthetic domains, Domain-Specific Batch Normalization (DSBN) was incorporated into both standalone and hybrid models. Experimental results demonstrate that the proposed approach achieves an overall accuracy of 94% in exact generator-level identification and 90.25% at the family level. The integration of DSBN significantly improves family-level classification performance, reaching 98.22%, highlighting its effectiveness in enhancing feature separability across generator families for medical deepfake detection.
Shahzad et al. (Fri,) studied this question.