Monday, 3 June 2019

Week 7 [03-09.06.2019] Can you tell which face is real?

Hi!
Recently, there was hot news about a paper entitled "A Style-Based Generator Architecture for Generative Adversarial Networks". The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis.

That research introduced a new, highly varied and high-quality dataset of artificial human faces.

The first approach to fake-face generation task was put together by Ian Goodfellow, the director of the machine learning at Apple’s Special Projects Group and a leader in the field when he proposed a new framework: Generative Adversarial Networks. In less than five years, all of that changed. Today’s AI-generated faces are full-colour, detailed images.

1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

You may have seen a website named ThisPersonDoesNotExist.com doing the rounds, which uses AI to generate startlingly realistic fake faces. There is also a website called WhichFaceIsReal.com, which lets you test your ability to distinguish AI-generated fakes from the genuine article.

2. Could you spot/are you aware of what kind of imperfection those generated faces have?

3. Do you consider any public security or privacy issues regarding those methods?

Articles and sources:

https://arxiv.org/abs/1406.2661
https://arxiv.org/pdf/1812.04948.pdf
https://www.youtube.com/watch?v=-cOYwZ2XcAc
https://www.theverge.com/2019/3/3/18244984/ai-generated-fake-which-face-is-real-test-stylegan
https://medium.com/@kcimc/how-to-recognize-fake-ai-generated-images-4d1f6f9a2842

35 comments:

  1. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

    Yes, I have heard about such images, both in the form of video and photos. I also know many possible applications. The first of these is an assistant to the elderly, who had the form of a virtual figure with artificial intelligence, who spoke to an older person and acted as her companion. The second application was putting under the actor an artificial silhouette of a known person, eg in order to record a false video in which a well-known politician speaks. The next application is of course all forms of entertainment such as games, virtual reality, movies, etc.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?

    Personally, I had the opportunity to interact with the effects that deviated from the actual character to a large extent and without difficulty I could recognize that it is a fictional character. However, this took place a few years ago and certainly it was just an example form, not the best existing one, also possible that there are much better solutions that would make it difficult for me to recognize the truth.

    3. Do you consider any public security or privacy issues regarding those methods?

    Of course, it is very important to create laws and regulations on this topic. It was frightening for me to meet for the first time with such a technique on the example of Putin making a strange speech that could attract many people and lead to conflict. I believe that technology can be used in a very bad and dangerous way, all the more so because the awareness of its existence among society is negligible. It is important to educate people and know about its existence, how to recognize it and defend it.

    ReplyDelete
    Replies
    1. Cezary, thank you for your comment. About the 3rd answer, I agree that it is important to educate people about the recognition of artificially generated content.

      Delete
  2. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    Yes, I’ve heard about such fake faces. I am impressed by the effects of the generative adversarial network. I couldn’t distinguish a real face from a fake one. Now I know how AI sees the human face :) I have no idea where such algorithms may be used – maybe during creating new things, like cars, for example.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?
    In general, I can’t see any imperfections. As for me, the faces created by GAN are indistinguishable from the real ones.

    3. Do you consider any public security or privacy issues regarding those methods?
    It seems to me that it will be a big problem. To date, official documents certifying someone's personal data include a face photo. Therefore, it can be used to impersonate non-existing persons. Facebook, Twitter etc. - which photos are authentic and which are not...? I think it's about creating an imaginary reality in which we stop believing our senses.

    ReplyDelete
    Replies
    1. Damian, I also have a problem to recognize artificially generated faces. Some clues gives an article https://medium.com/@kcimc/how-to-recognize-fake-ai-generated-images-4d1f6f9a2842 - you can try to look at imperfections of hair, background or teeth :)

      Delete
  3. Hi Artur,

    Thanks for relay interesting topic, you can find my comments below

    1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

    I heard about this technology. There are many places for that kind of application, for example Chat bots, I read the other day that there are algorithms that can animate human face movement from picture and do it In a meter that you wouldn’t even know.
    The second example I can think of can by video games, to push realism to next level

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?

    I am not aware of this spot, but during having fun and checking If am I as good as neural network, during learning process I was especially looking at light, some blur details, sense of realities :)

    3. Do you consider any public security or privacy issues regarding those methods?

    I think that the one of issue may lay on the training set – bunch of a photos that need to by push through a neural network that it can learned on it how to reproduce this images later on.

    ReplyDelete
    Replies
    1. Hi Klara! I also read about the rendered faces that could be animated based on a single photo (Mona Lisa was used as an example). Amazing stuff!

      Delete
  4. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

    Each of us who is interested in information technology and in particular image processing knows about such possibilities. Already a few years ago, an algorithm such as PatchMatch was written that enabled the installation of images where differences between them are blurred. In the same way, especially in the age of neural networks, artificial images of people were created.
    As for applications, it's primarily virtual assistants. The person talks more easily with someone who has a face. While conducting the discussion, we look at our interlocutor. Additionally, if he shows interest, we become happier.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?

    Yes, they are still not perfect. It's easy to recognize which is artificial and which is real. It's easy if we know that one's an artificial second real one. If we don't know that this counterfeit is already a little harder.

    3. Do you consider any public security or privacy issues regarding those methods?

    Any such method raises anxiety. There's no danger unless we realize it. Generating such a face gives you the opportunity to create an artificial and at the same time real identity. Today it's easy to take credit via the internet. You just need to provide a copy of the photo documents .... Hmm

    ReplyDelete
    Replies
    1. Michal, just as you've mentioned, the idea of a virtual assistant is a good example of application. PatchMatch was a breakthrough indeed.

      Delete
  5. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

    So I've heard about this project, at one time it was quite a popular topic on the Internet. As far as applications are concerned, I can think of the possibility of using this type of face as a character in a computer game. Of course, you can also use this type of collections to teach you how to recognize people through artificial neural nets.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?

    I've heard about the first page before but haven't seen http://www.whichfaceisreal.com. I took a test and made only two mistakes out of 10 times. So I think I can recognize them, although not all of them contain the imperfections that allow me to determine it.

    3. Do you consider any public security or privacy issues regarding those methods?

    I can't think of any dangers that might be associated with this technology. The only thing that is puzzling is the origin of the data collection from which this mechanism was learned. If our face were in such a collection, could it be a threat?

    ReplyDelete
    Replies
    1. Przemyslaw, 80% score is very high, congratulations :) Maybe when a person is a good in truth prediction, he/she is not concerned about such algorithms? :)

      Delete
  6. 1. Yes, I heard about artificially generated images of human faces. Such algorithms could be successfully used in computer games and augmented reality. Also in big-budget movies where a lot of extras are needed, an algorithm could support CG to produce whole crowds of people. Another application could be news or TV programs. Recently, I heard that news in Russia is being led by a robot that has been modeled on another television presenter. Maybe someone will use these algorithms one day to give Alexa or Siri a face.

    2. I went to the WhichFaceIsReal.com page you proposed and at the first attempt I did not guess which face is true, but with every subsequent attempt I managed to guess which face is generated by AI and which is a real photo. The generated faces often had some turbulence around various parts of the face, e.g. the ears or teeth. It looked as if someone tried to glue the face of several photos at once and forgot to clean some parts of the picture. It seemed very unnatural.

    3. There is always the possibility that someone will use the image of someone else for their own purposes in this way. The biggest danger I see is using the faces of famous and influential people for propaganda purposes or compromising them. Creating a non flattering photo like someone is drunk in a bad place. In addition, so-called catfish can enter a higher level if someone creates a fake account on community portals and scam other people by pretending to be a non-existent person.

    ReplyDelete
    Replies
    1. Hi Monika! I think that I would like to see my Home Assistant with a face :) However- yes, there is a large field for disturbances, harmful and lawbreaking.

      Delete
  7. Yes, I've heard about such algorithms before. A firth thing at the top of my mind is using it in film industry or in games to generate extras and characters. Generally it could be in used in media, especially in some embarrassing commercial spots or social campaigns where real actors could be reluctant to play a role in.

    I took a test on the web page you've shared. Indeed it was difficult to distinguish between fake and real pictures. On some there indeed were some distortions, but in general all looked very real.

    I'm not sure, I could think about some cases when people use generated faces to create ID documents etc., but it I don't know if it's any better to use generated face than real face of someones else in such criminal cases.

    ReplyDelete
    Replies
    1. That is true, fake ID with generated photos could be a real issue in the world of online verification methods.

      Delete
  8. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    With the help of neural networks it is possible to generate any number of photos of any human face - they look so realistic that we are not able to recognize that these photos are not real but generated with the use of appropriate algorithms. You can use them for super realistic characters in games, free extras in movies, fax news in news where "experts" and clones will speak on a given topic. Falsified information will reach heights. There will be a section in the police/corporations/military and secret services to fight/identify and use computer created characters and locations, there will be a mass of programs that will detect or create it all. Generally, it will be interesting,


    2. Could you spot/are you aware of what kind of imperfection those generated faces have?
    There is often an error related to the generation of teeth – incisors
    3. Do you consider any public security or privacy issues regarding those methods?
    At a certain level of fragmentation of features, we simply stop worrying. I might as well be concerned that today on the street an artist who remembered my face and unintentionally used some of its features in his next drawing, mixing them with the features of other people, passed me by. There is not even a very practical possibility to say that a feature in a drawing belongs to me because many other people in society have the same feature. Also the primitiveness is blurring.

    ReplyDelete
  9. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

    Yes, I heard. I also know that generated faces are using in campanies and comercials. I spent the last 5 minutes for looking for example of real-world example, but I couldn't find it! I'm sure that it was at least 2 example of commercial banner which contains artificially generated face.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?

    I use the page http://www.whichfaceisreal.com to check it on my own...I'm very suprised. I was wrong in most than half of all examples! Of course some contains distortions (I found strange/unnatural ear, and to be honest- too perfect proportions of some faces. All of we know that we aren't symmetrical). But in the heat of the moment, in first look I would say that every of that face is real! This generator are wonderful!

    3. Do you consider any public security or privacy issues regarding those methods?

    No...everyone could make you a photo on the street. If he had good equipment he could do it from long distance and you wouldn't know about it at all.

    ReplyDelete
    Replies
    1. Hello! Yes, the first contact with whichfaceisreal.com could be an awkward experience :)

      Delete
  10. Hi,
    Actually, I haven’t heard about artificially generated images which contain realistic human faces. It something new for me, both cool, and crawly issue. As a result, obviously I cannot propose too fast any extraordinary real-world applications of such algorithms, other than in films, TV shows, pc games, ads etc. creating alternative reality, just to fun.
    Definitely I cannot spot any imperfections of those faces. For me they are real ones, they could play as actors instead of humans. I find it a bit disturbing though, they all look like real people to me. The results are amazing and indistinguishable from the original images.
    As for last question, I cannot imagine any positive implementation of that regarding to security or privacy. On the contrary, I’d rather see mess possible to happen, if it was used to produce false important documents like identity cards, or even to present it in social media like facebook to cheat other people. In my opinion it can be helpful to criminals to pretend someone looking differently and this is dangerous for privacy and security. It will be interesting to see how the authentication of digital pictures will develop in order to keep up with technology.
    best
    marta

    ReplyDelete
  11. 1.Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    I have heard a bit about that. To do it we use GAIN network. Unfortunately I don’t know any application that used such algorithms to generate realistic faces.

    I found an article that explain how works generating fake realistic faces:
    https://www.lyrn.ai/2018/12/26/a-style-based-generator-architecture-for-generative-adversarial-networks/

    2.Could you spot/are you aware of what kind of imperfection those generated faces have?
    Maybe the positioning of the eyes is not realistic. I saw the faces generated automatically. For me they look like too perfect.

    3. Do you consider any public security or privacy issues regarding those methods?
    It’s important to protect our privacy. Generating a false face could be dangerous because people who want hide real face could use fake face to generate false profile. They could be log in to eg. Facebook and have account to bad things. I think people should be warned about potential risk using this technology.

    ReplyDelete
  12. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    Yes I have heard about GANs that generate faces. I think it may be used to create new characters in a movies that looks realistics. There will be no need to hire statists, if neural networks are cheaper.

    You may have seen a website named ThisPersonDoesNotExist.com doing the rounds, which uses AI to generate startlingly realistic fake faces. There is also a website called WhichFaceIsReal.com, which lets you test your ability to distinguish AI-generated fakes from the genuine article.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?
    I think that when they smile their teeths look unrealistic. The rest seems to be just bad quality photo and could be taken as a real perrson

    3. Do you consider any public security or privacy issues regarding those methods?
    Yes people may generate false documents and the it would be difficult to identify who is that if the photo is a fake. Or there may be some other types of cheatig with fake photos.

    ReplyDelete
    Replies
    1. Hello! Yes, those photos are near to perfection; I think that there is not as much to do to make them undistinguished from real photos.

      Delete
  13. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    Yes I am familiar with automated generation of human images. I think there were even TV transmissions that was based on such an approach.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?
    It is really hard or near impossible to say a difference between actual image and generated one.

    3. Do you consider any public security or privacy issues regarding those methods?
    Yes this could be an issue because such generated images could be used for phishing attacks and other security vulnerabilities

    ReplyDelete
  14. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

    Yes, of course I have. They can be used in games and augmented reality. I've seen also a couple of movies that weren't cartoons, but weren't feature films either and these artificially generated images were successfully implemented in them. I didn't like them, but still there were present there:).

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?

    As I already stated above, I didn't like the images containing human faces that I saw in a movie. There were imperfect, one could see straight away that they are artificial. Maybe, with time it will change. For the time being, I find those faces appalling.

    3. Do you consider any public security or privacy issues regarding those methods?

    I agree with the previous speakers. New technologies bring new risks and problems to the discussion. It's possible that one will unconsciously create an artificial face that really exists that is a face of a real person.

    ReplyDelete
    Replies
    1. I think that the technology would be improved and eventually a bring us a level of a real photo perfection very soon!

      Delete
  15. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?

    That's how I heard about this way of generating characters. The most popular known by all are probably characters from computer games. Movies have also been made on the basis of some games. A good example would certainly be "Avatar". It is true that these are not realistic human faces, but they were used for those that we can see in the movie.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?

    The given page of ThisPersonDoesNotExist.com generates very realistic faces using GAN, but you can notice the differences resulting from the technique used. Occasionally, you may see shortcomings, blurring, or in the case of people with glasses who have a problem with the face line.

    3. Do you consider any public security or privacy issues regarding those methods?

    Like any other type of artificial intelligence, GAN also generates faces based on training data. The question from where this data was received (perhaps from the fb.com portal or other similar)? Of course, everything is fine until anybody uses the tool in a wrong way, because later it will start to investigate. It would be good to introduce some legal regulations protecting our identity.

    ReplyDelete
  16. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    I've seen some material on artificially generated human faces on "2-minute papers" youtube channel, and also I think there was some news anchors in China state TV that was artificially generated?

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?
    It's all adversarial, also in a broader sense: each identified imperfection is fixed by researchers in next published paper. We could take resolution up to the levels single skin pores and capillaries are visible, but still it all can be generated in realistic way. So I think it's just temporary that we can sometimes tell apart generated ones from real faces.

    3. Do you consider any public security or privacy issues regarding those methods?
    In my opinion, we are past the point where one can believe all that is seen and heard, but with AI techniques, it is a matter of scale - new material can be generated in limitless amounts, even on demand. So I guess there are some new challenges.

    ReplyDelete
    Replies
    1. There was such an event! https://newatlas.com/china-ai-digital-news-anchor/57158/

      Delete
  17. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    Yes, I have heard about artificially generated images containing realistic human faces. I saw the site about which You mentioned in Your post – ThisPersonDoesNotExist.com. As You said, it pictures new faces generated by computers.
    As for some possible real – world applications, for example I think that it can be used in advertisements. In that case You wouldn’t have to hire some real actors to perform in it, which equals in some major savings for the company, which want to advertise its products.
    Moreover, it could be used in the news – from what I heard in China there is one news channel, where the anchorman is artificially generated.

    2. Could you spot/are you aware of what kind of imperfection those generated faces have?
    As artificially generated images are based on other photos of real human faces, they have a wide range of imperfactions coming from using the algorithm to generate the face – for example sometimes in this generated faces theeth are weird or other artefacts from photos, like oculars, which during the proccess of generating loose its holder.

    3. Do you consider any public security or privacy issues regarding those methods?
    Yes, of course. For example major problem may be seen in case of fake profiles on social media. Anybody could create a new profile using some artificially generated face, and then publish on such profile fake news or use it to create a lot of new profiles to give an impression that for example a politician has bigger public support than in reality.

    ReplyDelete
    Replies
    1. Hi! Yes, fake news in that context is very thrilling!

      Delete
  18. 1. Have you ever heard about artificially generated images which contain realistic human faces? Can you propose some possible real-world applications of such algorithms?
    Yes, I did hear about that kind of algorithms. I think that this kind of machine learning models are highly promising in many fields, first of all: entertainment. A highly developed GAN model can produce models that can replace human actors in movies, which would open completely new ways for movie producers: just imagine, it would be a tool allowing to have an actor looking exactly as the director wants him/her to look. Combined with advanced motion capture it would probably endanger the actor profession itself making it useless.
    2. Could you spot/are you aware of what kind of imperfection those generated faces have?
    I have read a couple of articles presenting the results of appliance of GANs and the way of revealing images generated by them. Those of older models, presented in 2015, typically have the following characteristic features, giving them out:
    the hairs are often either too straight, as if drawn with a ruler, or messy, without proper hairstyle;
    the faces can have some weird features, like beard and mustache on a woman's face;
    it is complicated for a model to maintain symmetry of the objects being on significant distance from each other, e.g.: earring at one side has completely different shape and size than the one at the opposite side, one shoulder can be unnaturally lower than the other, the eyes can be of different colors and shapes, directed in different directions;
    the background can contain some unnatural elements, the letters and digits are not readable. This effect is caused by the fact, that GAN models are trained with the use of the pictures, where the main part is occupied by human faces and background occupies too low percentage of the pictures, so a model simply cannot learn how to generate signs properly;
    the generated images can contain some regular noise.
    As for the newer models, they produce way more perfect images, where it is extremely hard to find the features revealing their artificial origin.
    3. Do you consider any public security or privacy issues regarding those methods?
    To my mind this technique can be potentially dangerous for public and private life: as for the first, imagine some major media resource being hacked and the news containing aggressive proclamations of some world leader are emitted. It would cause stock crisis, instability and many other dangerous things. As for the private life, on the other hand, almost everyone has an account in social media where he/she shares his/her photos and other private information, has friends etc. The modern methods of data analysis would allow, e.g. to generate a face basing on the pictures of the users, the account user likes and communicates with, causing him to trust to the person writing to him from this artificial face. This person would have higher influence on the account user in this case as he would trust him/her subconsciously.

    ReplyDelete