@Riccardo Perelli The image moderator model cannot be custom trained like the custom vision API offered by Azure cognitive services.
You can however create a custom list of custom images and then use the list to screen the images against this list. Basically, the API will then compare the image with this custom list and return a response. There is a limitation on the custom lists with only 5 image lists and each list limited to 10000 images. I hope this helps!!
If an answer is helpful, please click on or upvote which might help other community members reading this thread.