[nness]

There is already some interesting solutions posted here. If you wanted to try and tackle the issue with a stop-gap in the meantime you could add an image hashing step in the upload process to identify images that have already been flagged as NSFW or worse.

dHash is fairly simple to implement, and you might even be able to offload the hash checking at the database level. Comparing dHash's is just a matter of AND'ing the two hashes and counting the number of bits.

Obviously as the sample size increases so will the computation time. You could help the process by prioritising checks against new accounts, certain IP ranges (if you're seeing more or less content of a certain type from different countries or VPN providers) or if an account has a history of uploads in the past.

Its a horrible problem to have. Best of luck!