Enhanced Image Forensics: Unveiling Forgeries with Transfer Learning and YoloV5

This paper introduces an innovative approach to improving the detection of forged images by leveraging deep learning and transfer learning techniques, effectively identifying two common types of image manipulation simultaneously. Addressing the prevalent issue of malicious software forging digital images for disseminating fake information on social media platforms, the proposed approach simultaneously identify both image splicing and copy-move forgeries, thus reflecting real-life scenarios more accurately. The study utilizes the CASIAV2 dataset, comprising 4795 images, with 1701 authentic and the remainder forged. In our experiments, we utilized the CASIA 2.0 dataset, the same one used in. This dataset, consists of $\mathbf{1 2, 6 1 4}$ images, comprising 7,491 originals and 5,123 forgeries. The forgeries include 3,274 copy-move and 1,849 spliced images, all in JPEG, Deep learning methods like Convolutional Neural Networks (CNN) for image classification and YOLOv5 (You Only Look Once) for detecting forged regions are employed. Sequential, MobileNetV2, and Xception models are fine-tuned for the task. The experimental results demonstrate an impressive accuracy of $99.3 \%$. This research contributes to advancing digital image forgery detection underscores the potential of deep learning techniques in addressing complex real-world problems in digital media forensics.