Application Of Deep Learning For Image Deepfake Detector Using Convolutional Neural Network Algorithm

Abstract

Social media has long been used by the public in general as a means of exchanging information. Behind this commonly exchange of information, hide the malicious intent of those who are not responsible for spreading false information or hoaxes. This false information, which can come in various forms such as images, sounds, or videos, can actually be useful when used as stock photos or simply used as caricatures and satire. Unfortunately, false information often used on famous people instead to make them look like they said or did something that never happened. This certainly needs to be controlled, one of which is by using deepfake detector that aims to recognize false information pattern. Deepfake detector utilizes the computer's ability to self-learn to recognize that invisible patterns in images using one of deep learning algorithms, namely Convolutional Neural Network, which converts images into a collection of arrays containing numbers and then performs mathematical operations repeatedly on each layer. The result of the mathematical operation can then be used as a reference to determine whether an image is real or hoax. Author’s deepfake detector application using Convolutional Neural Network, specifically using the Resnet-50 model on hoax images created using AI with the ProGAN model, appears to be able to detect hoax images with the same model, with an accuracy of 85%, precision of 100%, and recall of 65%, but appears to experience decrease in accuracy when used in deepfakes with other models such as StyleGAN and BigGAN.