Autonomous driving is intended to revolutionise mobility. The car manufacturers are constantly reporting new successes in the field and are creating a spirit of optimism in the industry. The future of self-driving cars seems close, but technical problems continue to call safety into question.
One of these important and vulnerable technical components is image recognition AI, which is designed to scan and detect road signs and the like when driving autonomously. If artificial intelligence makes a mistake here and if a stop sign is considered only a speed limit, for example, this can become dangerous and lead to accidents.
Researchers from the University of Tübingen, in collaboration with the Max Planck Institute for Intelligent Systems, have demonstrated that even small stickers on street signs can irritate the image recognition and movement assessment of AI systems, reported “Heise Online”.
This pixel manipulation is hardly perceptible by the driver, but the researchers have known for some time. AI has no human idea of what a dog is and what a cat is. During training phases, she is taught to sort the pixel data and then decide which object it is.
Stickers or dirt on road signs can lead to misjudgments in the system in certain cases, which can also be exploited in a targeted manner. In a research project by Wieland Brendel, AI expert at the Bethge Lab at the University of Tübingen, and his colleagues, a targeted ai system was disrupted. In 85 percent of all cases, a stop sign was then confused with a speed limit, reports “Heise Online”.
Manipulation by outsiders cannot be ruled out
According to the researchers, not only image recognition, but also the movement assessment can then be restricted. The test results of the research algorithms have already been passed on to the automotive industry. However, this is not so easy to look into the maps and keeps its own algorithms secret.
Attacks on autonomous cars by outsiders are unlikely, but not impossible, according to the researchers. Certain patterns can irritate the pixel recognition of AI systems, putting the occupants of the car at risk, they warn.