SMILE III Safety analysis and verification/validation of MachIne LEarning based systems
|Coordinator||RISE Research Institutes of Sweden AB - RISE Mobilitet och System, Göteborg|
|Funding from Vinnova||SEK 6 128 303|
|Project duration||April 2020 - March 2022|
|Venture||Electronics, software and communication - FFI|
|End-of-project report||2019-05871eng.pdf (pdf, 6064 kB)|
Purpose and goal
New methods are required to ensure the quality of vehicle systems that rely on deep learning. SMILE3 built on the two previous SMILE projects by further developing the security mechanism "safety cage" for handling insecure input data. By supplementing perception systems with this mechanism, the aim was to demonstrate how one can argue that a machine learning-based system is functional. A further goal was to develop verification & validation of ML-based systems.
Expected results and effects
SMILE3 has developed a machine learning-based driver assistance system in the ESI Pro-SiVIC simulator, the emergency braking system SMIRK, freely available on GitHub as open source. For a limited operational environment, a complete safety argument for the machine learning component based on the AMLAS framework is available for further development. All data for training and testing of the models are offered freely. We expect that SMIRK will contribute to further research in ML and functional safety.
Planned approach and implementation
SMILE3 is a close collaboration of research and practical development. We developed SMIRK by following the iterative process in ISO / PAS 21448 Safety of the Intended Functionality (SOTIF). The safety argumentation followed AMLAS and was supported by workshops with all work packages represented. Collaborating on the open project SMIRK helped the project to explore relevant research issues despite the fact that the Covid19 pandemic largely forced work from home.