Agreed, this part got me wondering.

Thing is, not all request are as explicit such that you know what you are getting is "wrong"; therefor to the extend you rely on the moderation/censorship performed by the relay, this breaks in such an attack.

A handshake and/or periodical check should cover most attacks other than some super sophisticated targeted ones where networks themselves are pwnd, though im not sure about that.

I'd say those would be marginal enough to deem the measure sufficient