Deteksi Gambar Sintetis Buatan Ai Dari Berbagai Generator Menggunakan Gated Expert Convolutional Neural Network

Abstract

The creation of synthetic images has gained prominence in study as artificial intelligence (AI), especially texttoimages genartive model, continues to expand its capabilities. Vast amount of generators available, each having its own unique characteristic, giving the users a lot of choices to generate an images they want to create. While this advancement provides a breaktrough to generate images by using prompt, there is also a risk of its usages in crimes or other dishonest deeds. In order to mitigate that abuses of image generation technologies, scientist around the world tries to create a machine learning model that can identify an images created by a generative model. However, successful classification is challenging due to the diversity and complexity of AIgenerated images, caused by the sheer amount of generators having their own unique styles. In this paper, we suggest an approach that is based on Convolutional Neural Networks (CNNs) models to classify AIgenerated photos coming from multiple generators. One of the best approach is mixing the whole training data into one and train as a single entity. We propose gated convolutional neural network for better training efficiency and performance which slightly surpass the one trained on mixed dataset. By means of a thorough assessment and comparison with single CNN models, we demonstrate the advantages and disadvantages of several methods in improvising the classification of AIgenerated images.