A Trustworthy Explainable Deep Learning Framework for Copy-Move Forgery Detection Using Multi-Level XAI Techniques

Abstract

Copy-Move Forgery occurs when copies from different portions of images are placed at other locations to conceal or add content. This method is one of the most common and challenging types of digital manipulations. CNNs have shown great performance in detecting copy-move forgery; however, the issue with them lies in the fact that they are a “black-box.” Therefore, there is always suspicion about the results obtained by them in sensitive applications. We proposed an Explainable Artificial Intelligence (XAI) framework that uses our proposed detector alongside three well-known interpretability techniques: Grad-CAM, LIME, and SHAP. Our model not only correctly detects the forged region within the image but also offers understandable reasons behind the result obtained for the forgery location. Experiments on CASIA and CoMoFoD datasets show that we achieved 97.8%, 98.6%, 97.1%, and 97.8% accuracy, precision, recall, and F1-score, respectively, along with 0.72 IoU using Grad-CAM.