Thanks! Just had a meeting about my tech yesterday.

Google says the are using an "'ensemble model' trained using signals from multiple detectors." Sounds a lot like my PowerPoint I presented yesterday. They seem to be fishing for collaborators instead of offering their code to the public. I spent most of today working on developing one of the multiple detectors of my "ensemble model." Looking over the approaches of Google's collaborators, they seem to be trying to detect features that GANs can be trained to avoid, so unless the Jigsaw StyleGAN detector is particularly remarkable, even their entire "ensemble model" may not be difficult to co-opt/defeat. As I think I mentioned a month or so ago, StyleGAN improved a LOT over December 2019, so it is definitely a moving target that may well be able to keep pace with their detection efforts. I'm headed in a different direction to try to mitigate that risk.

Messages In This Thread