IMAGE FAKERY DETECTION BASED ON ILLUMINANT COLOR USING NEURAL NETWORK

• Main Author: Fauzan, Heri
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/43979
• Institution: Institut Teknologi Bandung

Images have taken the important role as a believable evidence to capture the reality. But, along with the era of technology, images are more easier to being faked by people. It makes we need a tool and research about how to define if an image is fake or real. Some tools like Izitru and Fotoforensics are already able to determine if an image has been modified by its digital structures but it never tell if the image being faked or not. It just a digital structures that might be modified by default, consciously or not. Also these tools using specific method that constrained by image format like JPEG and PNG. So we can’t use it to every image we found. In this research, author proposed to create an image fakery detection tool based on Illuminant Color using neural network so it will not constrained by any specific detail again. In this research, Illuminant Color method can be used by comparing two different ways to estimate an illuminant color in every pixel, there are Grey-World and Max-RGB. To use an neural network as a classifier, the result of Illuminant Color method should be extracted by feature extraction. Here, HOG edge was used to create a statistics of gradient over the image. This statistic result will take the role as input used by Neural Network’s architecture to create a learning model. Based on testing in this research, this tool can determine the image as fake or not. But the precision score using AUC-ROC tells that this tool only mapping the image very far away from the label value (0 as fake and 1 as real) with the 0.5 score. In different testing by manipulating the parameter value while creating some learning models, can be concluded that this method can be used to get good learning accuracy (more than 90%), but it depends on how the parameter (cell size, block size, number of hidden layer) is used.