AAI_2025_Capstone_Chronicles_Combined

19 0.20 intensity before dropping below 20% accuracy. On the other hand, the JPEG artifacts results tell a story completely different to those showed earlier. For the CNN, lower JPEG quality made the Real accuracy drop sharply at quality level of 10 while Fake accuracy stood roughly the same, but for the ViT, reducing JPEG quality makes the Fake accuracy drop down to 82% at quality level of 10, while the Real accuracy stays roughly the same. Figure16 Accuracy vs blur and image example, ViT

When comparing the accuracy against the blurred images, we also see some differences. While the Real accuracy in the CNN did not drop at all with blur intensity, the ViT loses accuracy down to 95% at the maximum blur level. As for the Fake accuracy, it drops similarly to that of the CNN, but only reaches a minimum of 65% accuracy at maximum blur, which is a much better result than the 25% Fake accuracy of the CNN at maximum blur. This makes it much harder to hide Fake images from the ViT

376