As ai generated images are being more realistic day by day, I am curios whether anyone knows a model that can distinguish whether an image is either fully/partially ai gen or real. If such a model exists, is it would be robust when detecting images generated by the models in the future.
I do not know of such a model. I do know there has been a lot of discussion on the topic. Especially in the context of deep fakes. One alternaive is to force AI producers to watermark. Speaking of which, last week, I discovered that chatGPT marks images it generates. I don’t know if does so across the whole platform, but it produced 2 marked ones for me. It occurred to me after posting a them, that chatGPT may track them.
Are you aware of what involvement AI companies have with an image after it’s been turned over to the client?
I think you would be interested in learning about Adversarial Neural Networks (ANNs), which have been used to do some generative tasks before. While I can’t say I know of any big detection models, the training methodology produces a detection model as a byproduct, which would be what you’re looking for more or less.
The model I am looking for should generalize well. If a model has training data consists of images generated by model A, B and C, I am expecting it can distinguish whether an image is ai generated even if generated by model D. It should generalize to unseen data.
I don’t think a model exists but you could build one up yourself probably. I’ll give you the instructions on how to do so.
First you need a database for your training images of both AI images and non AI images, after that you will need to label them yourself and convert the data of the images to numerical. Then you want to use a classifier on said images, likely KNN or naive bayes. This model should then learn on your images to attempt to find the AI images from the non AI images.
It will succeed if the AI images are clearly different from the non AI images and are in stable groups, it will fail if the model cannot tell where the classifier line should be.
Another supplement to this is to use an AI model to generate the images and then you can just have your actually-real-images database (so, all of the labeling is taken care of, at the cost of more computing power and potential skew/bias, depending on the strength and type of model you’re using).
use an AI model to generate the images and then you can just have your actually-real-images database
This strikes me as interesting in 2 respects.
-
As AI image generation evolves, detection systems would need updated to match pace. That’s how fraud detection works in banking and any sector. As crooks beat the system, the system adjusts.
-
AI trained to detect such subtle nuances is bound to lead researchers to identify key differences, and therefore direct them in what needs to be improved in order to match real image metrics.
That’s one of the cool things about developing with adversarial networks - they automatically are forced to learn new things, as they immediately start to notice a drop in accuracy with their models.
This is true - it’s the constant battle between the protagonist (image generator) and antagonist (image detector). This type of technology can, hopefully, enable a persistent competition between the two models until they are both so advanced that they can’t get any better (with current technology/computing power).