The CIFAKE dataset does unfortunately contain bias that allows the model without actually knowing the difference between real and synthetic images.

For example: Around 99% of real frog images aren't bright green, wheras around 97% of fake frog images are.