Explainable Deep Learning Framework for Accurate Detection and Interpretation of Copy-Move Forgery in Digital Images

Abstract

In the era of advanced digital image generation techniques such as copy-move forgery, the NDV is having increasing concerns regarding image authenticity particularly with the spread of digital images through social media, media and courts. This type of forgery, where a part of an image is replicated and pasted on the same image, to cover objects or materials poses severe challenges to the traditional approaches due to its inherent compactness in color, texture and noise. This research primarily focuses on the problem that deep learning models currently do not have sufficient interpretability and performance for the reliable detection and explanation of such forgeries under different image manipulations such as scaling, rotation and compression. This study proposes an explainable deep learning model which is capable of detecting copy-move forgery with high accuracy and is explainable as well from both feature and visual insights. The proposed technique has a hybrid CNN with both attention and feature matching modules, and an explainability module with Gradient-weighted Class Activation Mapping (Grad-CAM). The suggested framework has been evaluated on the benchmark datasets CoMoFoD and CASIA where accuracy, precision, recall and F1-scores were found to be 98.3, 97.6, 98.1 and 97.8 respectively which is higher than the other five state-of-the-art methods by 2-4 points. The approach was also able to withstand geometric transformations and compression and explainable by heatmap imaging. In conclusion, the proposed framework improves the detection and the interpretability of the deep learning-based copy-move forgery detection system, therefore it can be used in the domain of practical forensic research.