Such a scheme seems open to spam by design. I don't think there's any way to handle spam/moderation in a fully anonymous decentralised network.

Surely you could have small bayesian engines running in each client, and coordinate ratings through DHT. It would still have the issue of somebody potentially abusing such engine to effectively DDOS certain types of messages (e.g. mentioning a certain party), but you could try and fight that with modern consensus techniques from the cryptocurrency world.

If each client has a copy of the filter, then so does the spammer. The spammer can use that to craft spam the filter doesn't catch.

Then you need a system for distributing filter updates, and preventing the spammers from distributing bogus filter updates. Isn't this basically just the same problem as before?