Face recognition technology has been widely used in people’s production and life, but the verification of a large number of biological information can not help but cause people’s concern. What if the technology is cracked?
Recently, Yan Qi, CEO of Kuangshi, known as the unicorn of computer vision, appeared at the third Artificial Intelligence Conference. At the conference, Yinqi said:
“In the face of applications like deepfake, even if it has technical maturity and commercial value, Kuangshi will not apply it to products.”
On the one hand, this sentence emphasizes Kuangshi’s determination to protect biological information, but on the other hand, it confirms that deepfake technology has been very mature.
In fact, deepfake and anti face recognition and face recognition technology based on deepfake principle have emerged in endlessly. In 2019, it even exposed the news of Silicon Valley’s artificial Intelligent Company Kneron using DeepFake to successfully deceive Alipay and WeChat.
Sometimes just a pair of improved glasses is enough to cheat face recognition
With the increasing application of face recognition, is the face recognition era that we are entering really safe?
Deepfake principle based on deep learning
Deepfake technology comes from the emergence and development of GaN. Generally speaking, using deepfake to forge can be divided into three steps.
Step 1: provide face recognition data based on both sides
Taking song Xiaobao’s Zhang Ruoyun’s “baoniannian” as an example, the author of this video first imported Zhang Ruoyun’s image data in qingniannian, and then imported song Xiaobao’s image data, which completely restored the biometric features of the two aspects in the algorithm.
Step 2: deep learning
After the algorithm remembers their face information, it will start the deep learning mode. In particular, deepfake makes deep learning by constantly distorting the portrait and then restoring the algorithm. After repeated deep learning, the algorithm can restore any image with character characteristics, even if the restored object is not the original object (song Xiaobao).
The longer the deep learning time, the better the restoration effect
Step 3: output through different decoders
The last step is to distort the faces in the two videos, and let the algorithm through rich and deep learning restore the wrong video, that is, to restore Zhang Ruoyun in Qing Nian Nian to song Xiaobao, and song Xiaobao in the sketch to Zhang Ruoyun, and output the video. In this way, the video we get is what we see now in Bao Nian.
Since 2017, deep forgery technology has been active in the network. With the maturing of this technology algorithm, up to now, whether it is portrait or voice, video can be forged or synthesized, and can reach the degree that it can hardly distinguish the true from the false, and start to cause many social problems. Face recognition also faces the test of being cheated.
Rampant counterfeiting software and counter attack of face recognition manufacturers
With the full open source of deepfake and the research and development of face changing technology by algorithm manufacturers on the market, we can easily replace another person’s expression, action and posture in one video and achieve complete synchronization through simple app operation.
For example, deepnude software in the United States and Zao, a popular face changing software in China, were used before. In addition, in September last year, faceswap, an open source face changing tool, even made it to the GitHub list with more than 23000 stars.
On this basis, face changing software, face recognition masks, glasses and other products emerge one after another, which greatly disrupts the market order of face recognition. On this basis, many face recognition related enterprises have made their countermeasures.
Banned: Facebook, Amazon
In September 2019, Facebook announced cooperation with Microsoft, MIT and Amazon to jointly combat deep forgery. The program, known as the deepfake detection challenge (dfdc), aims to create open-source tools that can be used to train “anti-counterfeiting” models for companies, governments, media and other organizations to better detect whether videos have been tampered with.
It is worth mentioning that the team of China University of science and technology won the second place with 300000 US dollars
In December, Facebook’s artificial intelligence department announced the development of an advanced “anti identification” system that can identify the authenticity of the impact. In January, Facebook announced that it would apply the system to its software and remove all AI face changing videos from the platform.
However, the authenticity of forgery video is related to the time of machine deep learning and learning equipment. In fact, this method has been proved to be weak in the face of many deep learning or new face changing technologies.
Alibaba, Tencent: 3D face recognition + password verification
Companies like Alibaba and Tencent, which are involved in finance, are very vigilant against anti face recognition. Therefore, they generally use 3D face recognition + password authentication double insurance.
And this is also a common way to protect property security in the security industry.
Through the collection of specified actions, such as nodding, shaking head, blinking, mouth opening and so on, to achieve high-depth collection of face recognition data. At the same time, Ali also innovatively proposed a detection scheme based on light, shadow and position changes.
Jue’ao, senior algorithm expert of Alibaba security, said that this scheme is an integrated judgment of people, devices and algorithms. For example, there is a slight difference in the color of different mobile phone models. By changing the color and brightness of different areas of the face, the sensor can obtain more recognition information.
Jue’ao explained that the system will combine the changes of color, brightness, face posture, device space position and other conditions to judge the authenticity of the face. However, in the face of completely copied 3D face model, there is still a risk of breakthrough.
In addition, in the face of large consumption, Alipay will also start password verification and confirm the authenticity of payers through double insurance.
And Alipay’s face recognition algorithm comes from the open Face++.
Nearly 10 times higher than the common detection
At present, Kuangshi is the only enterprise in the industry with 1000 point level key detection capability, also known as the second generation key point.
What are the key points of the first generation? We call the former 81 points and 106 points as the first generation of key points. Their main task is to locate the more obvious and important feature points on the face. And 1000 points can not only locate the important feature points of face, but also accurately describe the contour of facial features. It’s exactly the difference between a point and a line. 1000 dense key points are the same as lines for facial features labeling.
This is also one of the core technologies of Kuangshi
It is understood that compared with the first generation of key points, QianDian can completely outline the shape of the face shape, eyebrows, eyes, nose, mouth and other parts of the face for more accurate face recognition.
According to the informant, at present, the open-minded face recognition has been in the forefront of the world. As early as 2019, the target detection algorithm of network attached storage (NAS) was completed. In the same year, the relevant research of Google just started.
In addition, Kuang Face++ worked with Alipay in 2015 to conduct in-depth face recognition training. As mentioned above, the effect of AI face changing is slightly different according to the time of deep learning. The longer the learning time is, the better the face changing effect is. After many years of face recognition training, Kuangshi face recognition algorithm has formed its own resolution ability, and constantly provides excellent face recognition services for users.
In fact, companies like Kuangshi, Alibaba and Tencent can completely counter the vast majority of anti face recognition software on the market. But why is there always news that face recognition has been broken in the market? The reason also lies in the pricing of its products.
Can’t afford or adult face recognition biggest pain point
According to a company employee in Guangxi, the biggest problem with face recognition systems on the market is that they can’t afford to buy them. The employee said that when she bought a 108 point face recognition system from a well-known CV company in China, the price offered by the other party was more than 300000 yuan, but it was difficult to grasp the quality of the enterprise system with a smaller name. After several times of selection, the final choice of a beauty software face recognition system.
In fact, there is a common phenomenon in the face recognition market: you can’t afford a good system, and you can’t look at a system that you can afford.
It is understood that the price of face recognition system is determined according to the number of key recognition points. For example, the price of 106 points is more expensive than that of 81 points. However, the cases of the above-mentioned kuangshiqiandian identification that can be applied in the market are relatively limited.
On this basis, a large number of enterprises, especially those involved in face modification, have joined the competition of face recognition, such as beauty software light face. The competition of face recognition market is more and more fierce.
It is undeniable that even if there are many software layers that interfere with the accuracy of face recognition, face recognition technology is still the apple of the eye under the wave of intelligence. With the continuous development of anti face recognition technology, face recognition technology will continue to update and improve in the future.
We also hope that with the continuous optimization of technology, more and more enterprises will be able to use and afford the relevant technology. After all, the best one is the one that can be used