Some time ago, 2 yuan could buy thousands of privacy-related face photos and topics rushed into hot searches. CCTV reported that on some online trading platforms, you can buy thousands of face photos for only two yuan, and the price tag of more than 5,000 face photos is less than 10 yuan, thus unveiling the "AI face change". The black production chain.
In the two cases of embezzlement of citizens’ personal information uncovered by the police this year, the suspects illegally obtained citizen photos through “AI face-changing technology” for certain preprocessing, and then generated dynamic videos through “photo activation” software, successfully deceiving people Face verification mechanism to commit crimes.
"AI face-changing" technology has been controversial since its infancy. On the one hand, the application of AI face-changing in the film and television entertainment industry has improved efficiency and effectiveness; on the other hand, the abuse of this technology may cause damage to property and reputation. At the beginning of 2020, Clearview AI's 3 billion face data was leaked, which caused many concerns.
Faced with a series of false face images and videos produced in batches using AI technology, it is a common method to use AI technology to automatically detect. However, the road is one foot high, and the magic is one foot high. The cat and mouse game of AI technology is constantly upgrading.
Source: Google AI data set
In 2019, Facebook, Microsoft, Amazon, Massachusetts Institute of Technology and other well-known companies and universities jointly initiated and held a face video deep forgery detection challenge. As of the end of March 31, 2020, no group of players has been able to detect a fake video system that has never been seen before in the black box data set with an accuracy of more than 70%.
In the GeekPwn 2020 false face AI recognition contest , which just ended in October, the challenge was further upgraded. The competition not only examines the ability of players to use AI technology to recognize false faces, but also introduces the concept of offense and defense, and studies how AI face swaps can create false face images and videos and bypass false face detection. At the same time, the use of AI technology for false face detection is improved. The accuracy rate.
is on the scene, and the players have to go through two stages of tests. In the first stage, the organizer provides pictures or videos that are a mixture of true and false. The false pictures or videos are generated using various popular algorithms, and players are required to use AI technology to identify true and false. In the second stage, each group of contestants made several videos and images on the spot, and after the release, each group identified the true and false.
Because these videos and images are very confusing, it is difficult to distinguish the true from the false with the naked eye. So after being prompted, it was discovered that the "Jay Chou" singing in the video was actually holding his face, and the host Jiang Changjian once exclaimed. As a celebrity, he has many public photos on the Internet and is easy to collect. When used for face changes, he has higher accuracy and is more difficult to recognize. Some contestants also said that when the picture or video is positive, only one photo can be used to change faces.
After fierce competition, the TSAIL team from Tsinghua University and Real AI finally won the championship with the highest defensive score in the game. This also means that their model for detecting fake images/videos is the most accurate among the five finalists.
Although there are only 20 minutes in the official game, the offline preliminaries have already started in August. Each group of players collected a large number of pictures and videos, and some teams even produced a data set of millions of pictures to train and test the model. At the same time, someone in each team is responsible for generating fake pictures and videos and studying how to bypass the detection mechanism. Finally, the generated pictures and videos are detected in the produced detection system, and this method is used to improve the defense capabilities of the detection system.
CAAD-Fake Face AI Recognition Competition Video
Common techniques for generating fake pictures or videos include DeepFakes, Face2Face, FaceSwap, DeepFaceLab, styleGAN, CycleGAN and other AI technologies or related variants. It mainly involves the process of face detection and recognition, feature extraction, determination of transformation matrix, and face replacement. Using different data and models, it is also possible to directly synthesize faces or videos that do not exist without the target identity as the basis.
source: https://www.thispersondoesnotexist.com/
false face detection starts from the principle and essence of face generation , based on the samples and models obtained by machine learning, combined with the image features of the face, to discover false people Forged details in the face. Like AI face-changing technology, the technology for false face detection is constantly updated and improved.
In October of this year, a paper published in IEEE PAMI ("Pattern Analysis and Machine Intelligence Transactions") claimed that biometric signals can be used to identify Deepfake videos with an accuracy rate of 97.29%, and it can also discover the generation behind the manufacturing of Deepfake model. Researchers believe that when the heart beats, it will drive the blood flow throughout the body. The flowing blood will produce subtle changes on the surface of the face, but the "person" shown in the fake video will not show a heartbeat similar to the person in the real video. mode. Therefore, it is possible to distinguish between true and false videos by detecting the subtle differences in the face caused by the heartbeat.
Source: https://ieeexplore.ieee.org/document/9141516
This seems to bring new ideas to false face detection. However, there is still some way to go from thinking to practice. The popularity of AI face-changing software has caused low thresholds and unlimited abuse, which also brings greater challenges to false face detection. Fortunately, there are many researchers like us who have never stopped on the road of offensive and defensive drills, early warning of risks, and promotion of the healthier and safer development of AI technology.
discusses
today. What kind of game do you think AI face-changing technology and false face detection technology will have? Nowadays, fake facial videos and pictures can be generated in large quantities, causing a lot of misleading and harm. How should we respond?
repost this article and leave your thoughts on AI face-changing technology and current situation in the comment area, or the impression left by CAAD false face AI recognition contest. As of 12:00 noon on November 13th, , readers with the highest number of likes will receive a GeekPwn 2020 Lei Feng package ; readers who rank second and third in in the number of likes will receive GeekPwn past commemorative sweater one Pieces.
