I think there's a lot of precedent for invisible watermarks for source identification in both copyright protection and exfiltration identification. Of course this is quite possible without compute intensive ML models.