Detectors of Synthetic Images: How They Work and Why They Fail

Detectors develop alongside generators. The task looks symmetrical, but in practice the defending side occupies the worse position, and here is why.

Approach 1. Artefact analysis

The detector looks for traces of generation: characteristic patterns in noise, unnatural frequency structure, repeating motifs. It works well against the models it was trained on and markedly worse against new ones. Each generation of generators devalues part of the accumulated evidence.

Approach 2. Provenance checking

Instead of analysing the picture, check its history. The camera signs a shot at the moment of capture, the editor records the changes made, and the chain is preserved in the file. The approach is more reliable but requires support at every stage: an absent signature proves nothing, because the overwhelming majority of files have none.

Approach 3. Common sense

Checking context: does a source exist, are there other angles, who published first, does the image square with known facts. In practice the most effective method, and the only one that does not go out of date.

Why detectors err

• False positives on heavily processed genuine photographs
• Misses on images that have been through recompression and screenshotting
• Complete helplessness against models released after the detector was trained

A race with no finish

Any public detector becomes a training aid for generators: developers check their output against it and adjust the model until the check stops firing. This is a built-in property of the situation rather than anyone's oversight.

The practical conclusion: do not treat a detector's verdict as proof. Neither a positive nor a negative result is a conclusion — it is one argument among others.

This material is intended for adults over 18 and is informational in nature.