For a long time, the "zamylivanie" and the pixelation of a part of the image containing private information, registration data, and persons very well helped to keep in secret what is necessary. But now, it seems, this method of camouflage can not be considered sufficiently reliable. The fact is that researchers from the University of Texas have developed a system of machine learning, which can with frightening accuracy identify smeared faces and text. Specialists, new system, told that it was not so difficult to do.
A person, looking at a blurred picture or a part of it, can not recognize what is depicted on it, but the neural network can very well and does it wonderfully, unmistakably recognizing text and pictures, smeared using different methods. The neural network was taught to "see" through pixelization and even learn that it is trying to hide the YouTube service with its proprietary blur tool. While the machine learning system does not know how to "blur" the picture, it can quite identify the object in the picture, comparing it with the original.
The researchers took the open software platform for machine learning Torch, algorithms for recognizing faces and text, connected everything and started learning the neural network. Accuracy of recognition was from 80 to 90 percent in the case of processed images on YouTube and 50-75 percent in the analysis carefully photographed with photo editors pictures. Worst of all, the neural network coped with pictures processed using the P3 (Privacy-Preserving Photo Sharing) tool - here the accuracy was only about 17 percent.
Worry about the fact that now everything that you have ever covered will become public, it's too early, but the results show that this day may not be so far away.
The article is based on materials .
Comments
Post a Comment