A photo breaks through the face recognition system: you can nod your head and open your mouth, netizens

Face recognition is hot again .

Just recently , CCTV has revealed a method to break through face recognition every minute ：

It only needs A picture That kind of .

△  Picture source ： CCTV Weibo

In the video demonstration, we can see , Any one of you , Use a paragraph that contains a nod 、 Shake your head 、 Driving video of speaking and other actions .

The characters in the original photo will do the same .

Although we know a static diagram , Now the probability is that face recognition cannot be unlocked .

But when it moves like this , The results are different .

therefore , The face recognition system can easily pass ：

△  Picture source ： CCTV Weibo

This video exposed by CCTV , It has successfully triggered a heated discussion among netizens .

Many netizens said that this way of breaking through the face recognition system “ terrible ”：

Let the picture move DeepFake

---

Although CCTV did not directly name the specific technologies involved this time .

But in terms of effect ,DeepFake You can do this .

DeepFake Everyone is familiar with , In short, there are two basic methods .

The first is to input a large number of facial photos of two people into the encoder , The encoder extracts the common facial features while compressing the image .

Then when restoring the image , Input the compressed picture of the first person into the decoder of the other person , produce “ In exchange for “ Facial effects .

The second is to generate a confrontation network （GAN）, Let two AI Algorithm （ Generators and discriminators ） Against each other .

The random noise is input by the generator and transformed into an image to be added to the real image , Judged by discriminator .

After a lot of cycling and training , Both have been improved , It can output realistic faces that do not exist .

△  Picture source ：3DCAT

But traditional DeepFake Need a lot of raw data , And it takes several days of training to achieve high-quality results .

If you want to achieve real-time results , What do I do ？

Li Hao （ you 're right , The one who killed Professor Matt ） The team proposed , take DeepFake And what he did before paGAN Come together , Made a new system .

So since , Without a lot of training data , This system can also render composite images in real time .

paGAN Make up for DeepFake The lack of a large amount of training data , Simply put, it is to put the workload of training on the stage .

There are three problems to overcome in real-time rendering ：

Need to process large amounts of data and use deeper networks to train better models , High resolution frames need to be generated and tasks can be arranged in parallel .

and paGAN After a lot of training in advance , I have analyzed the faces and expressions of many pictures . In this way, the internal data model can make changes when it comes to new graphics “ Conditioned reflex ”.

Plus paGAN New ones were used ML Methods and better underlying optimization , It achieves the effect of real-time rendering .

△  Picture source ：3DCAT