97%accuracy won the first prize of the national competition!Wuhan University Student Development Software Identification AI Face Change Technology

Jimu Journalist Xiang Yifan

Intern Ding Rong

When watching the video online, can you recognize whether the actor's face has been changed? Recently, a good news came from Hubei University of Technology. The undergraduate students of the School of Computer School, Cheng Yan, Cai Lei, Yin Xun, and Huangshan, who were undergraduate, and broke the technical difficulties under the guidance of the instructor Zhang Jixin. A "Video Stream DeepFake Detection System" with a face -to -face recognition rate of up to 97%has won the first prize of the 15th National University Student Information Security Contest.

Precise positioning face change features to help identify AI face changes

Although it is a work participating in this competition, the team's research and development of the "video stream DeepFake detection system" is not only to prepare for the competition. "Last year, we paid attention to some videos of AI face -to -face changes on the Internet. The original AI face -changing technology was developed, but unreasonable use will infringe on personal portrait and privacy." Team member Cheng Yan said that considering hope for hope AI face -changing technology can be properly applied. They decided to develop a detection system for AI face -changing in video images.

team member

In fact, before the team R & D software, there were relevant reports on cutting -edge academic papers, but most of them were not particularly accurate. According to Cheng Ye, the team compared other work discovery that most of the existing detection methods can determine whether the characters in the video have been changed through a single technical means such as morphological characteristics. In order to improve the accuracy, the team members decided to increase the technical level, not only considering the morphological characteristics, but also the four perspectives of the character's face gradient characteristics, spectral characteristics, and texture characteristics of the character to improve the accuracy of identification.

"In order to improve the accuracy, our detection software will integrate a variety of characteristics of face images to detect whether the characters in the video have been analyzed according to the detection situation." The team member Cai Lei introduced that due to the characters before and after the face change They are all unknown, so when they develop software, they do not find traces from the facial features of the characters in the video, but to directly find evidence of face -changing through technical means.

Take half a year to improve the accuracy of system detection

After the software development is good, in order to determine whether the detection software can accurately identify images that have been changed by AI, the team members have started a large number of tests. In addition to the online databases on the Internet, they will also download videos from various video platforms, test and detect the accuracy and adaptability of the software. Through experiments, the video stream DeepFake detection system developed by team members can not only identify high -resolution AI face -changing videos, but also have a high recognition rate for low resolution video.

"We have participated in the study and scientific research of the school's information security team in the first grade of college. In the process of software development, we also received a notice of the competition. In order to win the award in the competition, we started preparing for the competition for half a year in advance." Cheng Yan said, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the game, the competition, Earlier, every time the team members did an experiment, 50 to 100 sets of data were analyzed. When the performance indicators do not meet the expected results, the team members have been discussing the optimization plan and continuously verifying that the process alone took half a year. Through continuous trial and error, the video stream DeepFake detection system has a 97%accuracy rate.

"It is really happy to win the award, and our ability has been recognized. We have also learned a lot in the competition. Our professional knowledge reserves and professional literacy have been improved." After the game, the team four students said.

It is reported that the National College Student Information Security Competition is the highest level of professional college students in the field of information security in China. This competition has a total of 191 universities with 728 teams competing on the same stage. The team's "Video Stream DeepFake Detection System" achieved good results in the game. Another team of the school also won the third prize in the competition.

- END -