Archive notice: This article was originally published on June 14, 2019. Links and embedded videos are preserved as part of the historical record.
While various people shift the pixels in their own images into the “right position” and Canny AI has leading politicians sing “Imagine” together, some researchers are trying to address the ethical implications of the technologies. Adobe had already developed an AI tool last year intended to detect edited media created through cloning and removing objects or splicing – you can find more on that here.
Detecting Facial Manipulations in Photoshop
The new research project by Adobe and the University of Berkely now shows an AI-based method for detecting Photoshop manipulations made with the Liquify tool. By training a Convolutional Neural Network (CNN), the model can detect the altered facial features. For this, the researchers created an extensive collection with thousands of images from the internet.
Using a custom script, the researchers carried out automatic facial manipulations – with abstracted operations, for example to increase the width of the nose or reduce the distance between the eyes. Through random sampling, a training set of pairs of original images and realistically distorted modifications remained from this. In a subsequent study, the human eye detected the image manipulation in 53% of cases, while the trained neural network classified 99% correctly. The tool also identifies the areas and strength of the distortion. In the experiment, the tool also succeeded in restoring altered images to an approximation of their original state. The paper, as well as associated code (soon on GitHub) and further details about the implementation, can be found here.
Further information: Go to the paper’s official page
