BUFFALO, N.Y. — Deepfake photos and videos can pose as a real threat to our ever-growing digital world and expert Siwei Lyu believes these fake images will look even more real in the future.
“And if someone can manipulate that type of perception, it can potentially become very, very dangerous," he said.
Deepfake images are digitally generated images or videos made with artificial intelligence. AI can create high quality images of people who don't even exist, or create videos of public figures doing or saying things they never did.
But Siwei Lyu and fellow scientists at the University at Buffalo (UB) have created a new tool that can identify whether an image has been doctored. It analyzes light reflections in a subject's eyes. And according to new study, UB's deepfake spotting tool is 94% effective with portrait-like photos.
"We look into the eyes of the images. If it’s a real photograph, the two eyes see exactly the same things," Lyu said.
Lu says our corneas are very reflective, so they pop up in high resolution photos, but in a real photo, the reflection spots are in the same position on both eyes. In a deepfake photo, they are not.
“The model, being an algorithm, does not really understand the physical world," Lyu said. "So what happened is, it does not capture the simple, physical constraint that when two eyes look at the same thing, they should see the same thing.”
While these technologies continue to advance, Lyu says so will solutions to fix it, as well as the human ability to spot the fakes.
“We cannot totally eliminate them, just like we can’t totally eliminate the virus. But, we can coexist with them, so what we should do is always be cautious."
This story was originally published by Jeddy Johnson at WKBW.
