How does deepfake detection actually work?
Deepfakes got really cheap really fast. Face swap tools earlier used to need a powerful GPU and what now requires next to nothing. Same story with fake IDs, all you need is couple of prompts and you've got a driver's license, plus a thousand variations as needed
So for how detection actually works, beyond the usual "we use AI to catch AI." From what I've seen, and working in the industry, 3 things mainly:
- the face doesn't match the rest of the image: lighting, edges and blending around the face are slightly off
- the movement looks weird over time: blinks, head turns and expressions get glitchy between frames
- the face is too clean: it's missing the tiny skin pores and hair detail the rest of the frame still has
But the more tricky thing is doing it at scale, on every channel, without making the user experience a nightmare. And since a fraudster can retry a thousand times easily, even a tiny miss rate turns into a real headache.
This is pretty much my day job, im working at microblink so, i can be a lil bit biased. But we did score 100% on the IDNet deepfake benchmark, but I m not claiming any single check wins this on its own. It's more about stacking signals: document, face, liveness and behavior together.
If you’re also in this field, how are you seeing deepfakes show up in onboarding or support calls yet? And are you catching them or slipping yet?