Enhancing Image Manipulation Detection through Ensemble ELA and Transfer Learning Techniques

Abstract

Image manipulation techniques, such as copy-move, splicing, and removal methods, have become increasingly sophisticated, challenging the credibility of digital media. These techniques manipulate images at the pixel level, often leaving traces of tampering that can be detected through pixel-by-pixel analysis. This research introduces an innovative ensemble methodology that merges Error Level Analysis (ELA) with transfer learning leveraging deep convolutional neural networks (CNNs) to enhance image manipulation detection. The study involves extensive experimentation with various deep learning architectures and classifiers, with a focus on utilizing the CASIA1 and CASIA2 datasets for evaluation. The findings highlight that the combination of ResNet50V2 and ResNet101V2 models with Random Forest as the classifier exhibits superior performance compared to alternative ensemble techniques. This optimal configuration demonstrates high accuracy in discriminating between manipulated and unaltered images. The research emphasizes the significance of ensemble strategies in the realm of image manipulation detection, underscoring their potential for boosting detection accuracy and ensuring robust generalizability. The outcomes of this investigation shed light on the effectiveness of combining ELA and transfer learning for improved image authenticity assessment, providing valuable insights for advancing detection methodologies in the field. Here we achieved a promising outcomes, particularly with the Random Forest classifier, which attained accuracies of 97.671% and 92.497% on deep learning for the CASIA1 and CASIA2 datasets, respectively.