AAI_2025_Capstone_Chronicles_Combined

18

As it can be seen, blurring the image has the exact opposite effect on the accuracy than adding noise. We see how the accuracy on initially classified Fake images quickly drops, which clearly signals that the CNN was picking up the noise in the images resulting from the generation process to determine whether a image was real or not. This seems in line with what other research such as Yu et al. (2018) has found, where they note that GAN-generated images tend to have a distinctive noise fingerprint. Others, like Frank et al. (2020) have also exploited this fact for more advanced frequency analysis techniques also aimed to detect these noise fingerprints. It is interesting though, how for a image to be Fake this model would expect an overall good quality image (high AQI measurements) but that instead is more noisy. Now, let’s take a look at how these results compare with those of the ViT: Figure15 Accuracy relative to added Gaussian noise and JPEG compression, ViT

As it can be seen, these results look very different to those of the CNN. For starters, this model is much more robust to Gaussian noise. As a reminder, the CNN dropped the Real accuracy below 20% at 0.05 intensity, while this model resists up to

375