Combining Deep Learning and Traditional Methods for Effective Copy-Move Forgery Detection in Digital Images

Effective Copy-Move Forgery Detection (CMFD) in digital images tackles the increasing issue of image authenticity stemming from developments in image editing tools. The study devised a comprehensive approach that amalgamates Deep Learning (DL) with conventional techniques to improve the identification of copy-move fraud in digital photos, employing the CASIA v2.0 dataset for assessment. A deep Convolutional Neural Network (CNN) is employed for feature extraction, producing high-dimensional feature vectors that encapsulate the complex attributes of the images. The vectors are subsequently subdivided into training and testing sets, facilitating the training of multiple models, including VGG16 (Visual Geometry Group of the University of Oxford), AdaBoost, and Vision Transformers (ViT). Each model is assessed using measures like accuracy, precision, recall, and F1-score, enabling the identification of the most successful model for classifying photos as original or forged. Experimental findings indicated that AdaBoost attained the maximum accuracy at 99.50%, followed by ViT at 98.20%, and VGG16 at 89.30%. AdaBoost's enhanced efficacy, achieved by integrating the predictions of multiple base classifiers to improve accuracy, suggests its viability as the most dependable technique for CMFD.