Localization of Deepfake Facial Images Through U-NET Architecture

Traditional deep fake detection methods predominantly rely on binary classification, categorizing images as either real or fake, which often limits their transparency and interpretability for real-world applications. To address this, we propose a novel approach that integrates Histogram of Oriented Gradients (HOG) features into a U-Net architecture, enabling the capture of both fine-grained texture patterns and highlevel contextual information in deep fake regions. For enhanced interpretability, Class Activation Maps (CAM) are transformed into the HSV color space, facilitating intuitive color-based analysis. Contour detection is then employed to localize manipulated regions, and bounding boxes are generated using Dlib to precisely identify these features. By iterating through 68 facial landmarks, textual descriptions are generated to highlight alterations in facial features, providing a detailed understanding of the modifications. To further quantify the extent of manipulation, the resulting feature map undergoes a quantitative assessment by calculating the percentage of contour-detected areas. This method offers a comprehensive framework for identifying and analyzing subtle differences between genuine and modified images, advancing the transparency and accuracy of deep fake detection.