Hacker News Reader

Very naive approach but wouldn't training a model on a few specific buckets like "someone is naked", "image is explicit" and "there is a child" in the picture, would do the job without having to train explicitely on CSAM? If the model returns both a high probability of "there is a child" and a high probability of any other bucket than this image is classified as CSAM. It'd be high recall,low precision but one would be on a rather safe side.

This is already done, there's hashed indexes of known CSAM and you can apply vector steering to the latent space to shift detections to trigger on CSAM material. You can also use two detectors for age estimation and a separate one for nudity and use them in combination.

I believe this is how many CSAM classifiers work. There isn't a simple legal avenue to host CSAM for classifier training, for good reason.