Browsing by Author "Muhammad Awais Dar"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Unmasking gan variants
    (UMT.Lahore, 2025) Muhammad Awais Dar; Manahil Suriya; Muhammad Ali Malik; Muhammad Sajjad
    The authenticity of medical images created using a computer is now a problem due to the quick application of generative adversarial networks (GANs) in image creation. In this research study, we were trying to solve the problem of determining whether the type of GAN used in the generation of fake breast cancer ultrasound images was DCGAN, Style GAN, or Stack GAN. We were looking to create a system that would accurately recognize real and fake pictures and then classify fake pictures based on the source: DCGAN, Style GAN, or Stack GAN. We created 1000 images for every GAN odel from a real data set of 300 cancer ultrasound images. 300 images were randomly chosen as fake images from every GAN to balance the dataset. Real, DCGAN, Style GAN, and Stack GAN were the four balanced classes of the final dataset. Six deep learning classifiers, i.e., VGG16, esNet50, EfficientNetB0, and MobileNetV2 (custom-trained and pretrained), were trained. Class-wise report of performance, confusion matrices, and accuracy were used to benchmark all models. It was found in our experiments that the pretrained MobileNetV2 classifier outperformed all the others with a test accuracy of ~100% (99±1%). It illustrates how well transfer learning can find extremely fine distinctions between artificially created and actual medical images. Our system also performed favorably in labeling fake images correctly to their generative GAN model. The work proposed is the foundation for ensuring computer- assisted diagnosis systems and a sound method of identifying fake images within the healthcare field