The emergence of AI technology has promoted the popularity of face recognition technology, but is it really reliable to change face with AI? Maybe it may not be true.
Everyone here, have you tried to unlock your phone? Have you tried brushing your face for shopping? Have you ever played face recognition for animal expression games? If there is, then we must reach a consensus: AI has become more and more proficient in the face of people.
If recognition is just the first thing AI does on a human face, what is the second thing? From all kinds of signs, there is only one answer, that is, to change faces.
Of course, AI doesn't really give people a facelift (at least not at the moment), all it can do is change faces in the video. For example, the recent small screen-level video may have been seen by everyone.
The heroine in the video (exactly the face of the heroine) is Gail Plus, the actor of Wonder Woman. But of course, this is not a shameful little movie that he has starred in. Instead, someone used deep learning techniques to replace Gail Jiaduo's face with the body of the original heroine. At first glance, there is basically no flaw, but in fact, the two people are not very similar.
(Left is the original film role, the difference between the two is not small)
This ID is called deepfakes. It is always dedicated to sharing its star-changing video made with AI technology on Reddit. Almost all popular Hollywood actresses have been "concocted" by him.
Do you feel a little excited about your classmates? In the future, I want to see which star's film is done by myself. I can even replace my face with a play, and all kinds of YY can come true.
However, if the face of your relatives and friends is replaced? What if you change the face of the suspect shot at the crime scene to you? If you don’t know, the criminals send you a family kidnapping video with your face?
When we can't believe our eyes, the weight of all kinds of chaos and sin will definitely be greater than the little "welfare" that is illegal.
The horror of changing faces is that AI is very simple.In fact, it is not surprising that the video is changed. The familiar ones may be in the "Speed ​​and Passion 7" in order to let Paul Walker return, the film uses the avatar + dynamic capture + CGI animation, with Paul's two younger brothers as substitutes, and finally let the audience's old friends return On the big screen.
(Paul Walker and the two younger brothers who eventually became his substitute)
Although such a face-changing technique is superb, it does not worry us. Because this is the artistic need of the film, it must cost a lot of labor and post-effect costs, and requires professional video editing technology to achieve. This means that this precise face change does not appear in everyday life, nor does it pose a threat to ordinary people.
But AI is breaking down all this.
Let's go back to the deepfakes that made the actress's face-changing movie. This buddy is not only an old driver, but also a "technical live Lei Feng" who loves to share. He not only released his results for free, but also shared his own tutorials on making face-changing videos, as well as his own deep learning code and related data sets... Probably he meant, don’t ask me who’s the video. Let's do it yourself...
Of course, this buddy is not focused on female stars. The above is the tutorial he shared about how to replace Nicholas Cage with Trump. Perhaps he is deeply immersed in China’s “Expression Packâ€.
According to his sharing, making a star face-changing video is very simple. For example, in the video of Gale Gardot, he will first collect videos and pictures from various angles of Gail Gato on Google, YouTube and various network maps. Form a material library that can satisfy deep learning tasks for face replacement.
He then learns and understands the facial features, contours, movements, and mouth styles of the heroine in the original small movie through the machine vision related model provided on TensorFlow. Then let the model find the pictures and videos in various angles and various expressions under the various expressions in the material library, and replace the original video.
Although you can see that the video he made is still flawed in many details, not natural enough. But in general, it can be mixed, and the production effect is increasing.
The real problem hidden here is that using the open source AI architecture to change the face of the video is not too complicated and too avant-garde, but too simple and too easy!
In order to verify this, I sent the video and code shared by deepfakes to three friends who are engaged in machine vision related work. The feedback is basically the same: this thing is not technically difficult. As long as the basic functions of TensorFlow will be used, the computer graphics card will not be too bad, and it can be done in just a few hours.
Even people who don't even have the basics of programming can follow the tutorial step by step and collect enough material to make a face-changing video.
Imagine that when an enemy around you wants to frame you, just collect your photos and selfies and combine them with any sinful or even dirty video and spread it in your social circle. Everywhere, what is the scene that is chilling? It's like guns can be arbitrarily traded without review, unregulated, and at a low price.
After the underlying technology of machine vision development is getting better and better, video face-changing will inevitably continue to enhance its popularity at three levels:
1. Nearly no threshold to use. Face-changing data sets, source code, and architecture can be easily found today as long as you have the heart. As the technology matures, this trend will only intensify.
2. Can be tooled. Since the technology is not complicated, this function is highly likely to be instrumentalized. That is to say, the criminals can make it into an application. After purchasing, as long as the video is added as required and the image of the person is desired to be replaced, the face-changing video can be automatically generated to achieve a true no-threshold.
3. Deceptiveness is constantly increasing: Some related AI practitioners believe that the video of deepfakes has only undergone a preliminary learning and replacement process, and has achieved high completion without repairing and detailing. Then, if you further modify the anti-generation network, you can probably generate a video that is difficult to distinguish.
In short, when we know that the photo can be PS, the video is no longer trusted. And, not just video.
Rain is coming: the next stop is live + face changeAt the beginning of last year, the relevant team at the University of Nuremberg in Germany released an application, also known as Face2Face. The ability of this app is to track the face through the camera so that people in the video can follow you.
Due to its precise capture and real-time capabilities, Face2Face has caused an uproar since its birth. In its demo video, countless netizens questioned that this technology would become an accomplice to cyber fraud and kidnapping for ransomware. It would be horrible to question if the other end of the video call is not the one you know.
Of course, Face2Face is currently in a closed state, and users can only play the role they provide. But after more than a year of development, the face capture and replacement technology in live broadcast has also been greatly improved. Now we can see the background and props replaced in real time on the live platform, and using AI to replace the face in the live broadcast is already close at hand.
Coincidentally, AI's voiceprint recognition and sound synthesis technology is also advancing rapidly. For example, Adobe has released new sound synthesis technologies in the past two years. It is not a difficult thing for ordinary people to use AI to perform Conan's conversion with a bow.
With AI, changing faces and changing sounds in the live stream are moving forward in sync. So what is the impact?
Double-headed people broadcast? Trump is sitting in the White House office with you? A certain red meat is singing "Conquest" in the live broadcast? No problem, all can be.
Are you very happy? Of course, you may be happy with the live platform, but the small meat is not happy.
Think about it from another angle, if the same technology is used in video calls? If you receive a video call from a relative/friend, take your privacy or borrow money from you, you will find that it is a stranger who has forged it. If one can completely disguise himself as another person, will anyone be happy?
When we turned on the mobile computer, we found that everything was not true. It’s a very maddening thing.
It is not difficult for AI to change faces. Due to the existence of various application scenarios and super high entertainment, it is difficult for us to stop it. So what really hurts us is probably the legal problems and ethical traps hidden in it.
Basically, it can be very reliable to say that many live broadcast and video platforms at home and abroad are developing live-changing face-changing technology. And some solutions are quite mature. Imagine, after the face of the red goddess and small fresh meat, all night to open a live broadcast said some cater to the psychology of the words, the gift is not much to squeeze the platform? - Even if the user knows it is fake.
Of course, the formal live broadcast platform probably doesn't dare to do this, and using this technology will be very restrained. But what if a third-party plugin can do this? Or on a lack of regulated underground live/semi-underground broadcast platform? After all, interest and hunting can drive people to do all sorts of things. Once the threshold of technology is lifted, the legal problems that come in are likely to break.
The hidden ethical trap here is that the portrait right may be more complicated than ever. Whether it is a star or an ordinary person, probably do not want to be broadcasted by others as "easy tolerant" into their own way.
But the question is, how do you prove that ta is easy for you? Or how do you prove that you are you? We know that portrait rights are images and videos that you take yourself. But is the AI ​​model built with your facial data still in your portrait right?
Even more difficult is that you have no way to prove that the portrait model built by AI is directly related to you. After all, deep learning training is done on the invisible back end, and the producer can be said to be delusional or built by someone who looks like you. Or just one more than your face, is it not you?
There are still many more complicated ethical situations. For example, does a person have the right to portrait of a deceased relative? If a person wants to use AI to revisit a deceased relative and make a video call with the deceased, but another relative insists that it is illegal, then who should I listen to?
This is still the basic ethical and legal contradiction. Beyond this, there are a lot of illegal activities that can be done with AI face change. Such as fraud, extortion, fall, and so on.
Micro camera is the product of modern high-tech, also known as micro monitor, which has the characteristics of small size, powerful function and good concealment.
Micro cameras are widely used, suitable for aviation, commerce, media, enterprises and institutions, families and other industries. The emergence of miniature cameras brings convenience to people's lives, and at the same time, some phenomena related to corporate secrets and personal privacy also arise.
Bottle Camera,Spy Hidden Camera,Micro camera
Jingjiang Gisen Technology Co.,Ltd , https://www.gisentech.com