Project duration: 01.07.2019 - 30.06.2022 | Website: ki-absicherung-projekt.de

The use of artificial intelligence is key to highly automated driving. In AI Safeguarding, AI and safety experts from industry and academia have developed a methodology for a safety argumentation that systematically identifies, makes measurable and mitigates vulnerabilities of AI functions. The goal of the project was to reach an industrial consensus for a methodical safeguarding of AI functions for the use case of pedestrian detection.

Safeguarding functions that use AI-based algorithms is crucial for the German automotive industry in international competition. A consortium of OEMs, suppliers, technology providers and scientific institutions has developed an "industry consensus" on a methodology to identify and systematically mitigate inherent vulnerabilities in AI functions. The methodology includes a systematic approach to derive a stringent evidence-based security rationale. 

Merantix Momentum developed new methods in AI Safeguarding that help overcome known functional shortcomings and security concerns in deep neural networks. Results include a new approach to active learning, a novel approach to network compression through pruning, and the provision of a testing framework that helps gather evidence for a general security strategy for AI functions.