It isn't certain that you'd be able to tell WHICH input that the attacker used, at least not with my scheme where you hide who's using what input. Revealing who's using what input might not be optimal if a user want to use inputs already tied to himself AND some inputs that aren't already, and doesn't want the unlinked ones to become linked to him.

Also, it's hard to convince other that the attacker maliciously broke the scheme where none of the participants has any long-term pseudonym, thus it's hard to blacklist that one input. Also, a single input can STILL be used many times simultaneously, even if not many times in a row.

I don't have enough experience with programming, and absolutely not of crypto implementations, to be able to make a proper proof-of-concept of this, sorry. I'm hoping that somebody else *who does* will find this interesting enough to implement.

So far I'm thinking that my two-round version SMPC scheme with initial PoW before starting SMPC would be pretty decent, where you'd remember which pseudonyms that mixing has worked with in the past. Participants that don't supply PoW is ignored, and to stop a sybil attack where many nodes don't do PoW to try to make you calculate PoW for nothing you'd connect to multiple dozens of random nodes at once.
Edit: If the SMPC was given bad inputs, it would also reveal who it was that gave it the bad inputs (if it were to finish, but giving bad inputs + interupting the SMPC is just wasteful from an attacker's point of view).

And by the way, I'm one of those who wants the system to be as solid as possible in advance. Nobody deploys bad encryption algorithms and try to patch the algorithm as flaws is found. Not that I think that the other suggestions is bad, but I don't want people to get screwed over because somebody forgot to consider a simple attack that a little bit more analysis would have revealed and maybe shown how to stop.