International Journal of Science, Engineering and TechnologyAuthors: Chetan Khalal, Vaibhav Chavle, Prof. Amol Bhadange, Diya Gundecha, Shivam Dabade
Abstract: Deepfakes generated using Generative Adversarial Networks (GANS) have become increasingly realistic, creating risks related to misinformation, privacy loss, and digital security. Existing deepfake detection systems often rely on deep leaming black-box models that lack interpretability and struggle to generalize across different datasets or newer GAN architectures. To address these limitations, this paper presents an interpretable and physiologically grounded deepfake detection method based on iris and pupil analysis-features that GANs still fail to reproduce accurately. The proposed system uses a two-level detection framework. The first level evaluates pupil shape consistency through segmentation, contour extraction, and ellipse fitting. The Boundary Intersection over Union (BloU) score identifies irregular or distorted pupil shapes commonly found in GAN-generated faces. The second level performs iris gradient similarity analysis by generating Sobel-based gradient maps of both irises. The similarity score between the left and right iris gradient maps captures mismatched textures, contours, and reflections- physiological cues that remain consistent in real eyes but not in synthetic ones. Evaluations conducted on FFHQ realimages and StyleGAN3 fake images show strong results. While iris gradient analysis alone achieves an AUC of 0.94, combining both levels yields 97.9% accuracy, 96.0% precision, and an AUC of 0.979. This demonstrates that integrating biometric features provides a reliable, explainable, and computationally. efficient solution for deepfake detection.
DOI: http://doi.org/10.5281/zenodo.21719822
