Image forgery detection using deep learning

Amlee, Nasrul Fitri (2024) Image forgery detection using deep learning. Project Report. Melaka, Malaysia, Universiti Teknikal Malaysia Melaka. (Submitted)

Text (Full Text) Image forgery detection using deep learning.pdf - Submitted Version Download (1MB)

Abstract

As society becomes increasingly dependent on the internet, it also becomes more vulnerable to harmful threats. These threats are growing more vigorous and continuously evolving, distorting the authenticity of data transmitted online. Given our complete or partial reliance on this data, ensuring its authenticity is crucial. Images, in particular, can convey significantly more information than text, and we tend to trust what we see. Therefore, preserving and verifying the authenticity of images is essential. To address this need, image forgery detection techniques are expanding. Detecting forgeries in digital images is vital to restoring public trust in visual media. The objectives of this project are to segment tampered regions in an image, develop a high-accuracy model, and evaluate the model's performance. By using the CRISP-DM method, the project thoroughly identifies the problems and details the solutions. The result of this project is a trained model from a multi-modal fusion approach, called early fusion and late fusion, that can be employed in a web app to detect whether an image is real or fake and segment tampered regions in a fake image. Both the early and late fusion approaches achieved state-of-the-art performance in localization, with average F1 scores of 0.750 and 0.751 respectively across multiple datasets. For detection, our early fusion method demonstrated exceptional performance with an average AUC of 0.897 and balanced accuracy of 0.834. This will greatly benefit users on social media and authorities where this type of forgery is most prevalent.

Actions (login required)

View Item