Detection of AI-Generated Facial Images Using Convolutional Neural Networks

In the modern world, the artificial intelligence technology available enables the generation of human faces of which real world counterparts do not exist, and such potential offers a myriad of possibilities. Creativity can illustrate and fabricate work. Granted, technology of this nature can be wielded to serve the purpose of identity fraud, providing misinformation and other seemingly ‘immoral’ acts. Hence, this study aims to investigate the use of Convolutional Neural Networks (CNN) in composite face images created with ‘This person does not exist’ and ‘real life images’ download. Considering the study’s focus, the learning rate of 0.0001, sigmoid, 0.4 Dropout, and average pooling for tuning showed the desired learnt outcomes. The results were astounding, the model achieved 99% accuracy on validation and 97% accuracy on the training dataset. This accomplishment was attributed to a face’s underlying subtle features, such as its textures, symmetry, and visual interferences. Optimisation was conducted to measure generalisation, needing the model to perform on a new dataset with additional smartphone images. The accuracy was 84% for augmented and real images, with 5 outcomes correct of 6 sample images.