No, they use Magic Number among a bunch of various tricks. You can't really call it image recognition, more like 'data-hash comparison.' To train true image recognition would require a LARGE subset of stored illegal pornography.
https://en.m.wikipedia.org/wiki/List_of_file_signatures
Also check out "libmagic"
If you can do this (with different categories, for instance, determining what is nudity, then determining black people, to automatically classify porn with black actors automatically), it would be quite impressive. Results have nearly 5% error on CIFAR-10, which is full of clearly distinct categories - planes versus boats.
Not only impressive, but likely very profitable.
> Results have nearly 5% error on CIFAR-10
Any idea what the error rate is for human classification on CIFAR-10 images?