Hany Farid, a computer scientist at Dartmouth College in the United States, specializes in image and video counterfeiting. His clients, including universities, media organizations and courts, say image fraud has not only become more frequent but also more frequent The clearer he is, he tells Nature the 'arms race' between him and the counterfeiters you chase after me.
What is the first step in identifying false images?
An easy and effective way to do this is reverse image search. You place a search on Google Image Search or TinEye and they show where the image appeared elsewhere, a project at Columbia University Moving up to new heights, they began to look for parts of the target picture that changed from other pictures.
In general, if a picture is fake, we consider which patterns, geometries, colors, or structures will be destroyed, for example, if someone adds an object to a scene, we know that the shadows they added are generally wrong . In 2012, a video titled 'Golden Eagle Snatches Kid', one of my favorite examples, was found in 15 minutes and we found the shadows uncoordinated Where: The eagles and children are made of computer.
If the means of fraud is very hidden how to do?
We have a lot of analytics available. In a color photo, each pixel needs three values - one for each of the red, green and blue components - but most cameras record only one color per pixel, And fill in the gap by taking the average of the pixels around the pixel, which means that for any given color in a photo, each missing pixel has a specific association with its surrounding pixels if we add something or do something Modification, this association will be destroyed, we can detect it.
Another technique is called JPEG compression. Almost all pictures are stored in JPEG format, they lose some of the information stored, and each camera lost a lot of information in the storage If you use Photoshop to open the JPEG file, and then Save, and eventually there will be subtle differences from the original file, this is what we can detect, and I hope I can quickly identify the authenticity of any of your uploaded pictures; however, the current identification is still very difficult and need professional knowledge to find out Coordinated part.
Who uses your digital forensics service?
The organizations I serve include The Associated Press, Reuters, and the New York Times. There are a handful of professionals worldwide who specialize in digital forensics, so the scope of the work is limited, meaning you can only analyze really important pictures. However, Work is being done to broaden the scope of the analysis Last year, DARPA launched a large project I was involved in. They are trying to create a system over the next five years that will allow you to analyze tens of thousands of people a day This is an ambitious project.
For example, child pornography is illegal in the United States, but computer-generated child pornography is protected by the 'freedom of speech' provision of the First Amendment to the U.S. Constitution, and if someone is arrested, they may say The picture is not real, then I need to prove it, and I receive an email about the picture scam almost daily.
Your technology will be used in scientific papers?
I've been employed on several occasions at colleges and universities to investigate academic misconduct within schools, not long ago I went to the U.S. Research Integrity office where they asked me, 'How do we get automated tools?' In fact, we have not reached the level of automation yet. However, it is still possible to create a semi-automated process that detects dozens of photos instead of millions of photos per day, using tools such as clone detection to see if parts of the image were copied and pasted elsewhere. And colleagues are thinking about it, though small, but it is an important part of the DARPA project.
Please talk about fake video, right?
Researchers now have the ability to cut celebrity footage together to make videos that seem like they've never actually spoken, like some of Obama's videos. Researchers can also use machine learning techniques, especially to learn Create a GAN for the fake content and make a fake picture or short video so that a website that produces fake content competes with a 'classifier' site that attempts to authenticate and falsify the fake website quickly in the process .
I am very worried about the art of first-rate faking, and in five to ten years these technologies will reach first-class standards and after a certain period of time we will be able to generate realistic, audio-oriented world leader videos People are very upset and I would like to say that digital forensics technology lags behind video faking technology.
How to detect fraud video?
There is also a feature similar to JPEG compression in video, but harder to detect because video uses a more complex version, so machine learning techniques can be used to authenticate the video. However, the methods we used to identify video and identify The picture is similar in the way: Through observation, we found flaws in recorded videos lacking in computer-generated content. Computer-generated content is always too perfect. Therefore, one of the points to consider is that we can see the existence in the real world Statistical and geometric features?
The other technique comes from an outstanding study by William Freeman and his MIT colleagues: If you make a small change in your video, zoom in to see subtle changes in the color of your face that correspond to your pulse rate. Can distinguish between real and computer-generated people.
Machine learning algorithms can not learn to master these features?
In principle, it may be possible, but in practice, these algorithms have limited time and training data and are difficult to control which features the neural network uses to authenticate the video. GAN is only trying to trick it into training its classifier site. It will learn all the features that make it possible to distinguish the authenticity of a picture or video, nor does it guarantee that it can fool another classifier website.
My opponent must use all of the appraisal techniques I use to train the neural network to bypass the analysis of these aspects: for example, adding a pulse, which means that I have added difficulty to their work.
This is an 'arms race.' As we accelerate our rivals are also developing more sophisticated techniques to enhance audio, pictures and video, and the race will end only if the amateurs can not perfect the counterfeit. As you keep increasing the difficulty of making fakes, the more time and skill required to fake them, the greater the risk of being caught.
The original text of The scientist who spots fake videos was posted on the "Nature" press release on October 6, 2017