IMAGE FORGERY DETECTION OF SPLICED IMAGE CLASS IN INSTANT MESSAGING APPLICATIONS

• Main Author: THERESIA, MICHELLE
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/65922
• Institution: Institut Teknologi Bandung

As of 2021, smartphone user reached 79.84% of the world's population and 2.52 billion of them are active using instant messaging applications. As a result, the production and distribution of digital data exploded, and digital images were no exception. It also encourages the development of image manipulation techniques. On the other hand, image manipulation technology is also used to falsify information. This action is known as digital image forgery. There are several digital image forgery techniques and one of the most popular is image splicing. There are several methods for detecting spliced images and they are divided into traditional and deep learning. In recent publishing, proposed methods are mostly deep learning-based as they are able to learn features more generally. In this research, a modification is done on deep learning-based method for image splicing detection proposed by Meena & Tyagi (2021) in order for the method to detect compressed images from instant messaging applications by training compressed datasets through the Whatsapp application. The method consists of 3 stages, which are extraction of the input image’s noise residual using noiseprint model, feature extraction using ResNet-50, and classification, which is then applied to a desktop application. The solution is implemented using the Python programming language with some libraries: Tensorflow for noiseprint models, Keras for ResNet-50, PyCaret for classification, and TKinter for interfaces. The experiment done is to determine the classification model that has the highest accuracy using PyCaret library. From the result, it was found that the classification model with the highest level of accuracy is the Random Forest Classifier, which is 85.19%. However, the validation of the modified image splicing detection method using 100 DSO-1 datasets compressed via the WhatsApp application was unsuccessful because the accuracy was below the success criteria. On the other hand, the desktop application functionality is fulfilled and running well.