AAI_2025_Capstone_Chronicles_Combined

2

Introduction As generative AI models keep getting better, identifying whether an image is real or AI-generated is becoming a harder task. These fake images, also referred to sometimes as “deep fakes”, have been showing up more often in our daily lives, maybe as a product preview in an online retail or in posts in social media. Even though this technology has many legit entertainment uses, it is also being misused to perform different kinds of scams, like fake news or misleading product images (Mustak et al., 2022). This creates the need for creating new and more accurate ways of detecting whether an image is real or is generated by an AI. For example, a journalist may need to verify if a photograph is real before including it in an article, an online retailer may enforce that product images posted by its sellers are real to avoid deceiving customers, and a social media user may need to verify if an image in a post is real or not. To tackle this issue, in this project I trained two neural models to predict whether an input image is real or AI-generated. I used two of the most common architectures deployed for computer vision tasks, Convolutional Neural Networks (CNN) (Lecun et al., 1998) and Vision Transformers (ViT) (Dosovitskiy et al., 2020), and compared their strengths and weaknesses for this task. I also explored how different types of image artifacts commonly seen in photos on the internet, such as pixel noise and JPEG compression, affect the quality of their predictions. I used the “Deepfake and Real Images” dataset (Karki, 2021) from Kaggle, which contains a collection of real and generated images featuring people faces. The result from this research are two fully trained models with high accuracy detecting fake images, along with their limitations

359