^{Effectual Deep Learning Based Digital Image Forgery Detection Model Using Ameliorate CNN, VGG 16 and VGG 19 for Copy Move and Image Splicing Classification With Enhanced Performance}

Abstract

In this research, a pioneer passive approach for detecting and localizing tampering in images is proposed. In the past, image forgeries have created a huge impact in the world. Developments of various image processing technologies made it easier to tamper images with the objective to cover up or hide a significant feature in an image. In some situation, such as in a law-suit, it is necessary to prove that images are not tampered. Furthermore, in medical research, the validity of an image is critical, and this need will expand with the future standardization of data sharing among institutions or between patients and practitioners. Several more domains are reported in the literature that need image forgery detection with high accuracy; however, the majority of these works are done in high quality images with only one type of forgery. Many finer features are lost during compression from high to low quality images due to transformations such as smoothing, anti-aliasing, and moreover many techniques can be used for forging the images. Therefore, a new approach is presented to detect such forgeries with accuracy using Deep learning model architecture based on sequential CNN that can precede all previous attempts for images with both type of forgeries (copy paste and image splicing). Among all three evaluated models, Ameliorate CNN model demonstrated the strongest overall performance, achieving 93.11% accuracy, 95.54% precision, 90.63% recall, and 93.02% F1-score on the sampled batch of 833 images. Across the complete dataset of 9,501 images, Model 1 correctly classified 2,020 of 2,064 real images and 6,756 of 7,437 fake images, resulting in 97.87% real-image accuracy, 90.84% fake-image accuracy, and 92.37% overall accuracy. In comparison, Model 2 obtained the highest recall (92.20%) and 91.52% fake-image accuracy among the three models, its substantially lower real-image accuracy and precision reduced its overall performance. VGG 19 based model achieved real and fake-image accuracy comparatively more balanced as compare to VGG 16 based model; however, its overall accuracy and F1-score remained lower than Ameliorate CNN model. Overall, Ameliorate CNN model provided the highest overall accuracy, precision, and F1-score which demonstrate the strongest discrimination of authentic images, establishing it as the best-performing model among the three evaluated architectures under the experimental conditions of the study with improvement in enhancing feature representation and robustness in image forgery detection.