Researchers at Michigan State University are working with Facebook to find a better way to detect “deepfakes,” or images that have been altered to resemble a different person.
Deepfakes are becoming harder and harder to detect, and that could create big trouble in the hands of someone with malicious intent.
The research uses what’s called “model parsing,” in which an image is run through a network that’s able to detect unique patterns produced by a specific device through imperfections in the manufacturing process. It’s similar to technology that can identify which digital camera produced a photograph.
MSU researchers Xiaoming Liu and Vishal Asnani worked with researchers from Facebook to test the process on 100-thousand synthetic images. The result was substantially better at detecting deepfakes when compared to existing techniques, and that could mean more accurate investigation into coordinated disinformation in the future.