Deepfake Detectors


I recently came across a weird bit of news: Berlin’s ruling mayor Franziska Giffey received a call from Kyiv’s mayor Vitali Klitschko. But then she noticed that this was a fake video call. The counterpart was not the mayor but a synthesized image. She noticed this only after 15 minutes of conversation! This is both one of the more notable deepfake incidents and also a dangerous omen for the future. If a politician can be tricked about the identity of a colleague for this long, anything is possible!

So the idea for this post is to develop a commercially viable detector for video and voice call deepfakes, which can be plugged in to conferencing apps and which can alert the user automatically if someone is faking it.

I don’t want to go into the technical details of how this works, but there was a Kaggle competition with a million dollar prize for creating a deepfake detector.

So what about the business case? One main question is, whether to sell directly to users or just to the operators of conferencing software. The former would require that the software allows for components to be plugged in by the user. The cooperation of the video conferencing company would be required either way.

Training these ML models to detect fakes requires a lot of sample data from real deepfake attempts. Working with tele-conferencing providers would ensure that this data is available. The benefit for operators to claim that their users are safe from such attacks is considerable, so working with them might be easy.

Once detectors for deepfakes become a reality on the Internet, there will likely be a sort of arms race between the fakers and the detectors, with ever better techniques evolving. This would be expensive for the detector industry to keep up with, but the leading supplier of detectors could command a premium from their customers.