Abstract
As AI-generated images continue to close the gap between authentic and synthetic content, it becomes increasingly difficult to distinguish between the two with the naked eye. My project dives into the frequency domain to expose artifacts that aren’t visible in the spatial domain. I tested a simple MLP that only looks at a 1D FFT profile and compared it to a deeper CNN that processes the full 2D FFT magnitude spectrum. The difference between the two was noticeable: the MLP topped out at 59%, while the CNN reached 88%. This indicates that the frequency domain does contain useful artifacts, but only a model with sufficient depth and spatial awareness can effectively learn and utilize them.
Graphical Diagram
Graphical Abstract
Gitlab url
https://code.cs.earlham.edu/ctknight22/Senior-Capstone/-/tree/main