Current image-recognition systems, which are based on neural networks, are susceptible to targeted manipulation. This is what the researchers in MIT's LabSix team have found, focusing on Google's image-recognition system Inception V3. One of the objects they generated as part of their research was a 3D model of a turtle, which was unequivocally recognizable as a turtle to humans, but which the artificial intelligence (AI) behind Inception V3 classified as a rifle from almost every angle.
The AI initially classified a 2D image of a tabby cat as a bowl of guacamole; it was only once the image had been rotated slightly that the AI correctly classified it as a cat.
The fact that AI-based image recognition is sometimes unable to handle images from the real world at best leads to problems that require a degree of subsequent human rectification when it comes to standard applications like the automated archiving and cataloging of photos. In security-related areas, criminals may, however, exploit the weaknesses of these systems to cause considerable damage.