APPLYING TRUFOR IN DEEPFAKE IMAGE DETECTION TASK

Abstract

This study addresses the problem of detecting Deepfake images using the TruFor model, a forensic deep-learning architecture designed to analyze both semantic content and image-level manipulation traces. The methodology involves training and evaluating TruFor on a curated dataset of authentic and manipulated images, extracting noise residuals and computational artefacts to distinguish real from fake. Experimental results demonstrate that TruFor achieves high detection performance, with a pixel accuracy of 96.64% and a mean Average Precision (mAP) of 87.79%, indicating strong robustness and generalization across diverse manipulation types. These results confirm that combining semantic features with noise-based forensic cues significantly enhances Deepfake detection effectiveness.