Researchers from the University of California, Riverside have introduced a new technology for detecting changes.
The first branch distinguishes facial expressions (emotion) andpasses information about the areas that form this expression to the second branch - the encoder-decoder. The encoder-decoder architecture is responsible for detecting and localizing manipulations. As the authors of the work note, the solution they created not only detects interference, but also determines the areas of the image that have been corrected.
“Deep learning uses traits,detected by facial expression recognition systems to improve the performance of conventional change detection systems. This approach provides impressive results in the detection of facial manipulation,” says Ghazal Mazaheri, study leader.
Image: Mazaheri & Roy-Chowdhury, The UC Riverside
The researchers used to evaluatethe effectiveness of its model, two sets of videos, in one of which only emotions have undergone changes, in the other the acting character has changed. Experiments have shown that the developed system is much more efficient than existing analogues in terms of searching for changes in emotions and, at the same time, is not inferior to the most effective systems for detecting the substitution of the identity of the actor. During testing, the system detected 99% of fake videos.
Scientists believe that their development will help in the development of automated tools for detecting fake videos containing propaganda or disinformation.
Read more:
Look at the "silent" drone with a new generation of ion propulsion
Scientists figured out how to find dark matter using a conventional factory
Astronomers have studied 4 clusters of stars in the Milky Way: one of them is behaving strangely