Examining Deep Learning Models for Reliable Detection of Copy-Move Forgeries in Digital Images

The proliferation of digital image processing has led to an increase in copy-move forgery (CMF), necessitating effective detection techniques to ensure image authenticity. This study evaluates the performance of three deep learning architectures: Convolutional Neural Network (CNNs), Autoencoders, and Convolutional Neural Network - Long ShortTerm Memory (CNN-LSTM), using the Copy-Move Forgery Detection (CoMoFoD) dataset to identify CMF. The assessment parameters encompass accuracy, precision, recall, F1 score, and the duration required for computation. Among the various models discussed, the CNN-LSTM model achieved the highest performance, demonstrating an accuracy of 94.8%, precision of 92.4%, recall of 93.9%, and an F1 score of 93.1%. In comparison, the CNN and Autoencoder models achieved accuracy of 91.2% and 88.5%, respectively. Additionally, the CNN-LSTM model showed promising outcomes for identification following the implementation of the most common distortions; the identification accuracies reached 90.1% for rotation, 92.5% for scaling, and 88.7% for Gaussian blurring. The findings suggest that employing more intricate deep learning models, explicitly utilizing a single stream of convolutional neural networks, could enhance the accuracy of forgery detection. This lays the groundwork for practical applications in digital image authentication.