A new way to train deepfake detection algorithms improves their success

Deepfakes are images and videos which combine mixed source material to produce a synthetic result. Their use ranges from trivial to malicious, so methods to detect them are sought after, with the latest techniques often based on networks trained using pairs of original and synthesized images. A new method defies this convention by training algorithms using novel synthesized images created in a unique way. Known as self-blended images, these novel training data can demonstrably improve algorithms designed to spot deepfake images and video.

This post was originally published on this website.