AI-generated image detection using Machine Learning techniques

Abstrakt

Rapid developments in the field of generative
artificial intelligence have enabled the creation of photorealistic
images that are becoming increasingly difficult to distinguish
from real photographs. This work aims to implement or adapt
three contemporary methods for generated image detection: a
Photo-Response Non-uniformity (PRNU) extractor paired with a
custom CNN, an Error-Level Analysis (ELA) extractor coupled
with the same custom CNN, and an adaptation of the Latent
Reconstruction Error framework LaRE2. Each of the methods
has been thoroughly tested on a custom dataset, which was
constructed by combining part of the Tiny GenImage dataset
with images generated by a state-of-the-art transformer-diffusion
model named FLUX. The data, balanced evenly between real
and fake images, has been separated into five subsets, each
respective to one of the included models: ADM, BigGAN, FLUX,
Midjourney, and Stable Diffusion v1.5. Experiments were conducted
in three different categories, to ensure proper validation
of performance of the tested methods.