Researchers from the University of Washington, from the University of Michigan, from Stony Brook University and the University of California, Berkeley have discovered a way to hack self-driving cars by placing stickers on street signs.
Starting by analyzing the algorithm that the vision system uses to classify images, they used a series of different attacks to manipulate signs in order to trick machine learning models into misinterpreting them. For example, they printed stickers to deceive the vision system that an autonomous car would use to read a STOP sign as a 45 miles per hour, whose consequences would obviously be very bad in the real world.
In the article titled “Physical world attacks on machine learning models”, los autores demostraron cuatro maneras diferentes de interrumpir la forma en que las máquinas leen y clasifican estos signos utilizando sólo una impresora en color y una cámara. La parte más preocupante de estos experimentos es que todos ellos parecen muy sutiles para el ojo humano, camuflados como graffiti, arte o incorporados en la imagen de la señal.
Fountain: AutoBlog