Comment by happycry
4 days ago
Detection is one layer, not the whole solution. We've also built an AI model that inserts invisible watermarks for content provenance.
We've also been doing active work in source tracing: https://github.com/piotrkawa/audio-deepfake-source-tracing
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.