With $10 and a laptop, everyone can change his face

Since last year, face recognition has been banned in the United States. In May last year, San Francisco became the first city to prohibit the government from purchasing and using face recognition technology. This year, affected by the Freudian incident, IBM, an American technology giant, announced that it would officially give up face recognition technology. The next day, Amazon said it would stop providing face recognition technology to the police. In addition to the problem of racial discrimination, there is another serious hidden danger behind it. That is, people’s facial data may be used for “deep forgery.”. < p > < p > once upon a time, there was a certain threshold for deepfake. Although there are open source deepfake software and tools on the Internet, either the hardware requirements are too high, or the “fake” is not enough “deep”. All in all, it takes a lot of work for professionals to make a perfect deepfake video. < / P > < p > however, the worrisome deepfake is now “civilian”. The Scout even found an open source toolkit for deepfake for only $10 and a basic laptop. It takes only 24 hours for a face full of flesh and blood to become a handsome boy “Tom Cruise”. It is not too much to say that the degree of fraud is “stealing the sky and changing the sun”. < p > < p > in the video, a group of “technology houses” take video in their own laboratory, and use AI technology to change his face into the face of famous actor Tom Cruise < / P > < p > the process of changing faces takes five steps: video to picture, face extraction, training mode, face replacement, and composite video < / P > < p > first, they will do it for themselves Take a source video as material for AI to learn. Note that you need as many facial expressions as you can during the shooting so that you can better blend with the object you need to replace. < / P > < p > and then use video software to extract facial expressions. The software used here is AE video clip. When a person moves in the video, the outline and details of his face are “mattressed”, and pictures of different angles and expressions are generated. < / P > < p > in the video, we can see five parts: the most realistic photo is on the left, the face of the target character needs to be replaced in the middle, and the blurred part on the right is the preview of the final replacement result. < / P > < p > in the continuous learning process, the images generated by AI will be more and more close to the actual pictures, and the more real the final learning results, the higher the authenticity of the final “face changing” will be. It will also take longer. < / P > < p > although we can see in the video that the face of “Tom Cruise” is distorted because it is not matched well, this is also because AI has only learned for one day. But it’s still very realistic from a normal perspective. < / P > < p > although the video does not specify which deepfake tool these tech houses use. However, during the search process, he found an open source tool called “deepfacelab”, which is an open source deepfake tool released by a company named “Ctrl Shift face”. < / P > < p > and the famous deepfake video was made using “deepfacelab”. In this famous video, Bill Hardt unknowingly becomes Tom Cruise, and he looks like Tom Cruise when he was young. < / P > < p > at present, the toolkit has obtained more than 500 stars on GitHub, and it supports 64 bit pixel learning with minimum 2G video memory. In other words, you don’t need too much computing power to learn AI faces on your own computer. < / P > < p > the only installation requirement of deepfacelab is the corresponding version of the graphics card driver, even without CUDA and cudnn. And the software also has a Chinese official website and detailed Chinese installation instructions and tutorials. At the same time, it also brings a lot of face image data. < / P > < p > according to the tutorial, although there is no visual operation interface, the operation process is still simple and easy to understand. There is no code base of Xiaobai, as long as you carefully follow the operation steps, you can start quickly. < / P > < p > it is worth noting that although the basic requirements of this toolkit for computing power are very low, we have also mentioned that the higher the video memory and the longer the AI learning time, the higher the quality of deepfake video. Therefore, if you need to reach the above level of instant face change, I am afraid it will take a long time to run the expensive GPU in your computer. < / P > < p > nowadays, there are many tools and open source packages for deepfake. Many geeks and computer enthusiasts like to use different software and tools to make their own deepfake video or audio. However, don’t forget that the proper term “deepfake” is inherently deceptive. If abused, it will bring troubles to personal privacy, media credibility, judicial decisions, etc. < / P > < p > for example, as previously reported by Silicon Valley insight, in April 2018, the video of former US President Barack Obama saying “trump is a complete idiot” has been played more than 2 million times on twitter. In May 2019, a video of us Speaker Pelosi’s stuttering speech, which was forwarded and ridiculed by trump on twitter, caused a lot of attention, and eventually proved to be deep forgery The pseudo video generated by operation. < / P > < p > however, the face changing app “Zao” that appeared in China in 2018 was immediately banned from the shelves after three days of online popularity. Moreover, at the end of last year, laws and regulations were successively issued in China, which made it clear that deep learning, virtual reality and other new technologies should not be used to engage in activities prohibited by laws and administrative regulations. < / P > < p > in addition to these derived social impacts, deepfake also brings an insurmountable challenge to the face recognition technology itself. In 2018, s é Bastien Marcel and Pavel Korshunov, the proponents of the concept of “deepfake”, said that according to the current development level of face recognition technology, under the same model, the recognition error rate of face recognition with the same training time is as high as 95%. In other words, the existing advanced face recognition models and detection methods are at a loss in the face of deepfake. < / P > < p > the two scientists pointed out that with the continuous development of face changing technology, deepfake’s counterfeiting ability is also improving, and the two will “keep pace with each other” for a long time. < / P > < p > that is to say, if face recognition technology is compared to spider man, then deepfake is the “poison” attached to him. It will continue to increase with the enhancement of the power of noumenon. If it is used in the right way, it will become a great help; but if there are evil ideas in mind, its powerful power may even devour the noumenon. < / P > < p > so in the end, I want to remind you. Deepfake is a driving force for the development of face recognition technology, and can let some “technology houses” test the computing power of their computer GPU very well. However, if you use this technology to engage in crimes, it is putting the cart before the horse. Continue ReadingASMC, a lithography maker, was one of TSMC’s 14 top suppliers last year