We are all waiting for the moment when artificial intelligence, programs, technology learn to do practically the same, then we are. This is all quite real. But so far, few people think that all the same, this process will not be the same for machines and people.
For example, a group of researchers from Kyushu University discovered a way to deceive pattern recognition programs by changing a single pixel in an image. This pixel is positioned in such a way that it confuses the neural networks responsible for identification, and they begin to mistake dogs for cats and horses for cars.
The test images were limited to just a thousand pixels, but for higher resolutions and millions of pixels, the rule also works. Only it is necessary to change not one pixel, but several hundred.
The second group of researchers, this time from the Massachusetts Institute of Technology, managed to go even further. They were able to print a sculpture of a turtle using a 3D printer, which the program, even seeing from all sides and from any angle, took it for a rifle. This experiment poses a much bigger problem for image recognition technology, since the whole thing happened in reality, and the program was allowed to analyze the object from all angles, but it still did not determine its true shape.
Together, both studies show how robots and programs need to travel before they reach human level. After all, our brains cannot be fooled by replacing one or even a hundred pixels.
The only pity is that the sources do not provide an explanation of how it turns out to deceive the program. What do you think? What is the secret of "broken pixels" and how can you cheat with a 3D model?
Nikolay Kudryavtsev
Promotional video: