Together with TÜViT and the technology company ZF, the Federal Office for Information Security has started a new project that is intended to provide insights into how the IT security of AI systems in cars can be tested “according to generally accepted criteria”. In the medium term, test criteria for AI systems are to be developed from the results of the study, which will result in a modular technical guideline. The directive, in turn, can contribute to the development of future car safety reviews and international standardization efforts.
A prerequisite for optimally protecting the AI systems is knowing how attackers proceed, says project manager Dr. Georg Schneider from ZF’s AI Lab Saarbrücken. Therefore, one would simulate attacks in the laboratory and analyze the reaction of AI systems. Above all, so-called adversarial attacks on a current AI-based traffic sign recognition were examined in different variants.
With 120 instead of 60 things
These attacks involve manipulation of optical signals, which are then misinterpreted by the AI systems and to which they then react with unwanted operations. In road traffic, this would mean, for example, that a manipulated 60 km/h sign could be read as a 120 km/h sign.
In order to be able to benefit from the advantages of AI systems in the field of mobility, safe use is essential. So far, however, tools with which the risks can be checked as well as concepts and methods for this are missing, says the head of the cyber security department in digitization and for electronic identities at the BSI, Dr. Silke Bargstaedt-Franke. According to Bargstädt-Franke, the complexity of the use cases based on interactions between classic software, AI systems and the physical environment does not allow the development of test standards on a theoretical basis, one has to identify practical problems in operation.
The focus of the project is now the testing and further development of requirements, methods and tools. Targeted use cases will then be tested in two follow-up projects. The technical guideline then aimed at as a project goal could, for example, be used in international standardization efforts within the framework of the UNECE World Forum for Harmonization of Vehicle Regulations.
(ur)