A Decade-Old Method for Detecting DeepFakes

Back in 2011, there was a paper on a novel video processing method called Eulerian Video Magnification. A variety of papers on this topic is available at MIT labs for deeper perusal. In my reading, the two most significant papers are "Eulerian Video Magnification for Revealing Subtle Changes in the World" (Youtube video) and "Detecting Pulse from Head Motions in Video" (Youtube video).

Although I have not seen any advancements in this area since 2018, I firmly believe this technology is a promising tactic to discern real human video from both AI-generated avatars or highly manipulated videos.

To combat the OODA loop of deepfake advancements, we ironically will need an AI model to ingest and classify valid Eulerian rhythms for known celebrities and politicians. Going forward, I envision computational optimizations that will allow:

1. Social media and news outlets to overlay a "confidence of authenticity" rating over streaming videos and recorded press conferences.
2. Video conferencing adapters actively monitoring and reporting the "confidence of authenticity" on participants.

Further down the rabbit hole, we have yet to see a similar technology for voice only deepfakes - going beyond the non-verbal cues to actually detect and measure breathing and pulse rates.

My biggest fear, though, is that we do nothing - and leave it up to the general population to determine whether their news feeds are real or fake.