Seven tricks to catch a neural network before a lame | Trends

Seven tricks to catch a neural network before a lame | Trends

You do not know what to believe? Has it happened to the internet like Pedro's grandfather (and the wolf)? Normal. The expertise of artificial intelligence in the generation of false images is constantly increasing. How to trust again, after seeing a perfectly fake Obama stating that "Trump is a complete idiot"?

We can not stop progress in these techniques. But we can help you to detect when what you have in front of you is a false image painted by an algorithm that is thought to be very clever but does not finish understanding, for example, how the hair works. These are the traces of lying, according to Kyle McDonald, specialized artist in computing.

  1. The phantom lock and other unusual hair. In principle, the only scenario in which we would have an isolated lock in the middle of the face is the hairdressing salon. When you see hair that seems to come from nowhere, distrust. In addition, it is usual that the artificially generated manes have unreal textures, more typical of painting.
  2. Sinister denture. Overlapping teeth, irregular sizes, spaces without sense, textures too soft … "Neural networks can assemble a general scene, but still have difficulties with repetitive details, such as teeth," explains McDonald.
  3. Asymmetry too asymmetrical to be real. Do you see unpaired earrings? Is an ear missing? Does each eye have a different color or shape? THAT IMAGE LIES.
  4. The elements of the background look like a painting by Dalí. In these false photographs, the back can be especially revealing. Waves without sense, traces that aspire to be characters but do not correspond to any recognizable letter. "The network just replicates general textures rather than real scenarios."
  5. Strangely uniform fillings. At this point the idea is similar to that of a child who does not know how to color, but instead of leaving the outline of the drawing, the network fills the monochrome spaces with spots where "semi-regular" patterns can be detected. "Older networks have much more prominent noise patterns," adds McDonald.
  6. Rainbow Surprise. Eye to the white spaces. From the shirts to the eyeballs may appear iridescent shadows that show soft tones of color where they should not be.
  7. The visible invisible slope. In its attempt to copy the real models, the network may end up creating a lump indefinitely at the height of the area where a slope would have to appear.

Do you want to put what you have learned into practice? Here You have a small challenge You have five seconds to decide if the images are original or have been artificially generated.


Source link