Toshihiko Yamasaki and Kaede Shiohara at Japan's University of Tokyo (UTokyo) trained algorithms on self-blended images to better spot deepfake images and video.
Typical training involves pairing unmanipulated source images with doctored images, which limits detection to certain types of visual artifacts.
The UTokyo researchers used training sets comprised of synthesized images in order to control these artifacts, and to better train detection algorithms to find aberrations.
They found the modified datasets improved accurate detection rates by up to 12%, depending on the original dataset to which they were compared.
Yamasaki said the method works best on still images, but he envisions that "in the near future, this kind of research might work its way onto social media platforms and other service providers so that they can better flag potentially manipulated images with some kind of warning."
From University of Tokyo (Japan)
View Full Article
Abstracts Copyright © 2022 SmithBucklin, Washington, DC, USA
No entries found