back to article Can you tell real faces from fake AI-created ones? It's tough! Plus: Facebook's chief AI scientist talks hardware

It's Monday. It's a new week. The coffee's on. The hangover's over. Let's brighten your morning with some developments from the world of machine learning. More AI fakery: A seemingly growing number of academics, industry types, and policy wonks are wringing their hands off over the dangers and perils of fake content being …

  1. Anonymous Coward
    Anonymous Coward

    Shadows....

    From these two samples only, it's interesting that the fake faces are those without shadows. If that happens with other images, the AI should learn how to apply shadows to the faces as well, to make the images more realistic. It is true that you can use a lightning that will deliver shadosless portraits, but in most instances there will be shadows.

    1. katrinab Silver badge

      Re: Shadows....

      Also, if the photo has "unusual" features like different coloured eyes, that probably suggests it is real, as the so-called AI thing won't introduce rare/unusual features like this.

      1. TRT Silver badge

        Re: Shadows....

        The fake Air bnb also has a similar incidence of sunlight through all the windows, which would suggest a long, thin apartment with all the windows on the same side, yet the styles of the window frames vary greatly from room to room.

      2. Nick Kew Silver badge

        Re: Shadows....

        It might, if it's seen such examples in training.

        I've seen several individuals with startlingly mismatched eyes. Dogs, not humans. Perhaps because humans frown upon incest for ourselves but inflict monstrous in-breeding on dogs.

        1. TRT Silver badge

          Re: Shadows....

          Woah! I'll be sure to let David Bowie, Robert Downey Jr, Benedict Cumberbatch and Demi Moore know how you think they acquired their heterochromia. Well, I might need a ouija board for David Bowie, but ... man!

          1. William Towle
            Alien

            Re: Shadows....

            Pub quiz factoid: Bowie's left eye actually had its pupil dilated more that that of his right, apparently due to muscle damage after a teenage scrap. While this sometimes has the appearance of heterochromia (I thought so too*), that isn't what it was.

            * it was "anisocoria" (...TIL), which I imagine the computer won't reproduce in many of its images either.

      3. oiseau Silver badge
        Big Brother

        Re: Shadows....

        ... the so-called AI thing won't introduce rare/unusual features ...

        Probably ...

        Until it gets taught/learns how to do it.

        Just like us.

        O.

    2. 's water music Silver badge

      Re: Shadows....

      I spent a while looking at thispersondoesnotexist the other day. Leaving aside those images with a gross distortion artefact in the background I noticed that almost all had what looked like weird skin folds in the neck or around the edge of the face, sort of like wrinkles but not quite. They seemed to re-use the same teeth a lot which was another giveaway. I got bored before I got a fake/real guess wrong (ok it didn't take that long). Without the training a good set of fake faces and the expectation that exactly one in each pair is fake I would not be so confident

    3. Eddy Ito Silver badge

      Re: Shadows....

      Not all have wonky shadows. In the lower set it's clear one is lit from the side while the fake could be a head on flash. After playing a few rounds it becomes easier to tell what to look for. Many of the fakes have strange features like several hairs floating in space offset from the rest. Another with glasses was missing a temple piece and other such things that make you think "something's off". In the lower pic on the fake gent, what would be his right ear seems behind his head rather than attached to it while the few strands of the woman's hair down beside her neck is something that is natural but would unlikely to be included by an AI.

  2. Nick Kew Silver badge

    I see a hint of a shadow: on the fake woman's neck, and on the fake chap the right side is darker. Bear in mind that clear shadows come from sunlight or powerful artificial light, and won't exist in many perfectly-real photos.

    The fake woman at the top seemed obvious to me. The mix of mostly-young features with a definitely-old crowsfoot by the eye. But the fake chap, with no pretensions to be either young or old, has no such easy giveaway - to my eye, at least.

    1. Anonymous Coward
      Anonymous Coward

      "clear shadows come from sunlight or powerful artificial light"

      Shadows are a matter of size (relative to the subject) and direction of light, not intensity. Small light sources will create deeper, sharper shadows. Larger ones softer shadows. Frontal lightning will remove most shadows, as they will fall where the imaging sensor can't see them. But flat lightning is common in amateur flash snapshots and paparazzi images, while most photos will have some shadows (high-key images have little or no shadows but are a very different kind of lightning).

      IMHO, without an algorithm which includes lightning information, an AI will discard shadow information when it gathers faces features, and then will have to create an image without them. Anyway, it would probably need a 3D model of the face to apply shadows correctly, and probably this isn't generated, albeit an approximation may fool most viewers.

      If the fake images were compared with a flat-lightning real image, it would have been more difficult to spot them.

      Another hint is the lack of background, evident in the first photo, less in the second, where there's still a slight fall-off as the light comes from the left, while both fake photos have very simple backgrounds, as it's easier to synthesize them. Again, putting the image on a realistic background would make them harder to spot.

      The woman face is interesting, because it's not symmetric enough, and the lightning highlights that.

  3. macjules Silver badge

    Yann LeCun, Facebook’s AI veep and chief scientist

    Anything involving Facebook cries out for a subtly placed "t" at the end ..

    1. JassMan Silver badge
      Trollface

      Re: Yann LeCun, Facebook’s AI veep and chief scientist

      Being French as he is, a "t" at the end would be absolutely meaningless. Even if you change the "u" to an "o" which is the translation of your favoured word, it would still be pretty meaningless in this case as a con in French is someone slightly dim and usually gullible, something I rather doubt Yann would be, since his salary is likely to be well more that any randomly selected 20 elReg readers combined.

      Besides, didn't your mother ever tell you it is rude to attempt to make fun of anyone based on an attribute over which they have no control?

  4. Anonymous Coward
    Anonymous Coward

    Whatever your AI detection rule is, there will be a real world counter

    The problem is that we're getting closer to the algorithms we as humans use. Asymmetry is not as uncommon as you think - I recall a photography study where they mirrored facial halves to form a face, shadows and reflections can be calculated once the facial geometry is defined, even facial colour differences can have a natural origin.

    There is no absolute rule or certainty, it's a game of probabilities. You can probably tweak them upwards, but the idea that there is a definite, absolute, binary yes/no algorithm out there is a tad unrealistic.

  5. DavCrav Silver badge

    I seemed to be pretty good at 'Which Face Is Real?'

    Got the first one wrong, learned what the differences were, and next six all correct. Then I got bored.

    It's some form of human learning algorithm. Very powerful stuff, but difficult to explain to the layman. "You show humans a few images, and they seem to work it out for themselves."

    1. tcmonkey

      Re: I seemed to be pretty good at 'Which Face Is Real?'

      Same here. Did 10, got 9 of them right. Some of them were blindingly obvious, others not so much. Something in the blending gives it away, but it's hard to put a finger on (except in the cases where it's blatantly wrong).

  6. Anonymous Coward
    Anonymous Coward

    I seemed to be pretty good at 'Which Face Is Real?'

    I saw "Alita Battle Angel" at the cinema this weekend, and came fairly rapidly to the conclusion that "Alita" sat in the uncanny valley for me *despite* the best attempts of the SFX crew. If only I had a spare £100 or so to afford the (anglicised) original manga... For some reason I find them more immersive than a movie.

    Saying that, I must be getting old, I'd had enough of watching cyborgs being torn to pieces before I'd got half way through the film.

  7. Khaptain Silver badge

    I hate the bastards doing this.

    It only takes a couple of seconds to realise how this is going to end up creating more problems than solutions.

    A little bit of NVidia magic, a dabble of Facebook profiling, a little bit of magic fairy powdering snapchat/instagram powder and you are suddenly an internet meme with all the worst possible unthinkable habits, and it's all on video ......whether you want it or not....

    This isn't even Tinfoil hat imaginary scenarios, this is now and it's real....

  8. Ken Hagan Gold badge

    So what's new?

    This new story appears to say no more than "We trained a computer on several million real faces and now it can draw something that resembles them.". I see no *new* threat to society here.

    If I want a photo that looks like a human being doing something of my choice, there is this pre-existing technology called "acting" that has sufficed for over a century. If I want a photo that looks like a particular person doing something, we have technology for wrapping one person's face around another person's head. It has been around for a decade or two. It started off really expensive and is now relatively cheap, but as a society we seem to be coping. Perhaps it is the case that for cases where it really matters whether the video is real or not, there are ways to corroborate things. Perhaps people are slowly wising up to the fact that "the camera never lies, except nearly all of the time".

    1. wtrmute

      Re: So what's new?

      https://youtu.be/cQ54GDm1eL0

      This is a video of Jordan Peele have a fake Barack Obama call Donald Trump a naughty name. The news is that the technology for doing this is becoming a consumer-level tech so we can all make our political enemies do the worst immoralities our fevered minds can conjure up...

      It's a sick, sick world and we're neck deep in it, and sinking by the week...

  9. N2 Silver badge

    Perhaps

    Theyve looked at the Daily Mail? plenty of fake faces to be seen there.

  10. I.Geller Bronze badge

    Natural Language

    As far as I know, the computer can only understand numbers, signs, and words. That is, to understand an image it must be described by numbers, signs and words. Means everything that is in the computer falls under Natural Language.

  11. Michael Wojcik Silver badge

    Traffic signals

    If the sensors can’t detect a green light, human drivers have to take over

    Pah. This is Boston. Just count the number of cars illegally passing you on the sidewalk - that will tell you when it's appropriate (it's never safe) to go.

    That said, autonomous vehicles won't succeed in Boston until they have a configurable contempt level, like Scud the Disposable Assassin.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019