Global Relational Probing in Vision Transformers: Approach to Detect AI Generated Images

The increasing use of a generative AI like diffusion models has now set up this ongoing struggle that will not easily end between those generating fake images and those who try to identifying them. The very first techniques tried to find technical errors, like finding odd frequency patterns or leftover artifacts from the generation. These worked fine on older AI but once they are applied to newer generators, the methods seem not to work anymore. This review paper looks at the latest work with AI for detecting fake images and divides people’s methods into two basic types. One side of researchers continue to focus on artifacts and hope to get better at catching tiny technical mistakes. The alternative group is going towards looking out for what are called semantic oddities. The middle approach is where something like usual Vision Transformer gets it done just by itself. This proposed prototype is named Global Relational Probing (GRP). The main idea is easy to get; an usual ViT gets tuned, and then the self-attention layers look for important high-level semantic mismatches. To check if this technique actually is effective, workflow was set up that is easy to follow by other people. All of it runs with open source programs and the dataset collections anyone can use, also relies on explainable AI so everybody can see what system decided and why. GRP is built to survive new kinds of threats, can work with many kinds of the image generators.