Using AI to Conceal Confidential Information in Images

image provided by pixabay

This post is also available in: עברית (Hebrew)

Researchers from Japan, China, and Finland created a system that replaces confidential parts of images with visually similar alternatives using GenAI. The researchers created this system to provide a more visually cohesive option for image censoring and help preserve the narrative of the image while protecting privacy.

Associate Professor Koji Yatani from the Graduate School of Engineering at the University of Tokyo, who developed the system, explains: “We found that the existing image privacy protection techniques are not necessarily able to hide information while maintaining image aesthetics. The resulting images can sometimes appear unnatural or jarring. We considered this a demotivating factor for people who might otherwise consider applying privacy protection… So, we decided to explore how we can achieve both—that is, robust privacy protection and image useability—at the same time by incorporating the latest generative AI technology.”

According to Techxplore, the researchers then created the “generative content replacement” (GCR) computer system, which identifies what could be considered a privacy threat and automatically replaces it with a realistic artificially created substitute (like replacing personal information on a card with illegible letters or replacing a private building with other landscape features).

Yatani explained that there are currently several commonly used image protection methods, like blurring, color filling or simply removing the relevant part of the image. “Compared to these, our results show that generative content replacement can better maintain the story of the original images and higher visual harmony.” He added that in testing, they found that participants couldn’t detect GCR in 60% of images.

The GCR system currently still needs a lot of computation resources, so it is not available on any personal devices yet. Furthermore, while the original system was fully automatic, the team added an option for users to customize the final outcome a little more.

While some could be concerned about the risks of this type of realistic image alteration, the team seems to be positive about its advantages, with Yatani saying: “For public users, we believe that the greatest benefit of this research is providing a new option for image privacy protection.”