Project Details
Home/Project Details
QuishGuard
Project description
Detecting the images that were never photographed
The Challenge
Generated images are now good enough that a person cannot reliably tell, and the detectors that can are usually answering a narrower question than the one being asked of them. "Is this AI-generated" and "was this photograph manipulated" are different problems with different failure modes.
A detector that is right about yesterday's generators and wrong about today's is worse than a stated uncertainty, because it gets trusted. So the work has to be evaluated per generator and per manipulation type rather than reported as one accuracy figure.
The Solution
Treat it as a multimodal problem: the pixel evidence and the metadata and compression history disagree in informative ways, and a model that sees both has more to work with than one that sees either. Evaluation is broken out by attack type, so the report says where the detector is weak rather than only how good it is on average.
Skills and deliverables
Features delivered
The hard part
Generalisation. Every model in this space risks learning the fingerprint of the generators in its training set, which is why the evaluation includes generators the model has not seen and why the result is reported honestly rather than as one number.
The Impact
A research build with the model, the training pipeline and the per-attack evaluation in the repository. It is an experimental project by design and the record is labelled that way — the contribution is the evaluation discipline as much as the detector.
