A Comparative Analysis of Image Forgery Detection Techniques

An image is an essential medium for information sharing. Growth in technology led to the development of various applications like Adobe Photoshop, Snapseed, Gimp, and Prisma. These applications make it very easy for anyone to modify an image, giving rise to a new field of research called image forgery. Different pixel-level, block-based, camera-based, and mathematical algorithm-based techniques were developed to identify forged or modified images. Deep learning methods showed promising results as they considered more features than traditional methods. The study aims to conduct a comparative examination of image forgery methods. This paper also proposes an eight-layer CNN-based model with a relu activation function for the first seven layers and a sigmoid for the last one. Pretrained models VGG-16 and VGG-19 are used for comparison with the proposed model as they are the most widely used image classification models. In the proposed approach, the image dataset is first used to create a new dataset by applying the ELA (error level analysis) method to the original images. It is used for training the proposed model. The Casia dataset and MICC F2000 datasets were used for experimentation. This paper presents a comparative analysis of forgery methods based on the results. Results reveal that the proposed CNN model trained on ELA images outperforms other pre-trained models. All the models produced better results for MICC-F2000 than for the Casia dataset. The proposed model showed promising results on the small-sized image.