Deep Convolutional Networks For Copy-Move Image Forgery Detection

The authenticity of images is increasingly compromised, making them unreliable as evidence in critical applications. Common forgery techniques include copy-move, where a portion of an image is duplicated and repositioned within the same image, and splicing, which merges elements from multiple images to create a falsified version. This study introduces an efficient forgery detection framework that combines Scale-Invariant Feature Transform (SIFT) with a Convolutional Neural Network (CNN) to detect copy-move forgeries effectively. The proposed approach is evaluated using the MICC-F2000 benchmark dataset, comprising 2,000 images, of which 1,300 are authentic and 700 are forged. The CNN model achieved the highest test accuracy (99%), outperforming ResNet-18 (87.14%), hybrid CNN+SIFT (77.14%), and a 1D Autoencoder (55%). The CNN’s streamlined architecture of two convolutional layers with max pooling and dropout (0.5) proved optimal for detecting localized tampering artifacts, while deeper models like ResNet-18 struggled with over-parameterization. Interpretability analysis via LIME confirmed the CNN’s focus on semantically relevant regions, aligning accuracy with transparency. These findings emphasize the efficacy of lightweight architectures in forensic tasks, challenging assumptions that complexity guarantees superior performance.