[bpcrd]

We've actually deployed to several Tier 1 banks and large enterprises already for various use-cases (verification, fraud detection, threat intelligence, etc.). The feedback that we've gotten so far is that our technology is high accuracy and a useful signal.

In terms of how our technology works, our research team has trained multiple detection models to look for specific visual and audio artifacts that the major generative models leave behind. These artifacts aren't perceptible to the human eye / ear, but they are actually very detectable to computer vision and audio models.

Each of these expert models gets combined into an ensemble system that weighs all the individual model outputs to reach a final conclusion.

We've got a rigorous process of collecting data from new generators, benchmarking them, and retraining our models when necessary. Often retrains aren't needed though, since our accuracy seems to transfer well across a given deepfake technique. So even if new diffusion or autoregressive models come out, for example, the artifacts tend to be similar and are still caught by our models.

I will say that our models are most heavily benchmarked on convincing audio/video/image impersonations of humans. While we can return results for items outside that scope, we've tended to focus training and benchmarking on human impersonations since that's typically the most dangerous risk for businesses.

So that's a caveat to keep in mind if you decide to try out our Developer Free Plan.

[Eisenstein]

What's the lead time between new generators and a new detection model? What about novel generators that are never made public?

I think the most likely outcome of a criminal organization doing this is that they train a public architecture model from scratch on the material that they want to reproduce, and then use without telling anyone. Would your detector prevent this attack?

[coeneedell]

There are three observations that are helpful to know about here: A: High quality, battle tested architectures are sold via an API and samples are therefore easy to retrieve at scale. B: lower quality, novel architectures are often published on GitHub and can be scaled on budget compute resources. C: Often these models perform well at classifying content generated by architectures similar to those they were trained on, even if that architecture is not identical.

As for actual lead time associated with our actual strategy, that’s probably not something I can talk about publicly. I can say I’m working on making it happen faster.