Forgery Detection in Digital Media using Neural Networks

Abstract

The widespread availability of digital image editing tools has led to an increase in manipulated media, making advanced forgery detection techniques essential. This project presents a robust approach to identifying forged images by utilizing Python and a Convolutional Neural Network (CNN). The CNN acts as the core of the detection system, achieving remarkable accuracy with a training performance of 98% and a validation accuracy of 92%. These results demonstrate the model’s ability to effectively differentiate between authentic and tampered images. For this study, a dataset comprising 12,615 images was used, including 7,492 genuine images and 5,123 altered ones. This diverse dataset ensures a comprehensive assessment of the model's performance. To improve detection accuracy, the system integrates Error Level Analysis (ELA) as a preprocessing technique. Each image is resized to a standardized resolution of 256x256 pixels before applying ELA, which helps reveal inconsistencies in compression artifacts. Ideally, unedited images should display uniform compression, whereas discrepancies in compression levels may suggest potential alterations. The processed images are converted into NumPy arrays for further analysis. By integrating deep learning with CNNs and leveraging the subtle variations identified through ELA, the proposed system not only achieves high detection accuracy but also pinpoints areas within an image that may have been manipulated. Implemented using Python and a structured CNN framework, this project significantly enhances digital media forgery detection, with promising applications in fields requiring image authenticity verification.