10 - Safety and Security
ON THE APPLICABILITY OF SAFETY NETS: A SAFETY-BY-DESIGN SOLUTION FOR CERTIFYING NEURAL NETWORKS
J.M. Christensen¹, T. Stefani¹, E. Hoemann¹, F. Köster¹, S. Hallerbach¹; ¹German Aerospace Center, Germany
This work analyzes Safety Nets for certifiable AI in aviation, combining neural networks and lookup tables to guarantee correct runtime behavior. It evaluates architectural trade-offs, identifying compact designs achieving high accuracy with minimal storage. Results show significant size reductions and provide the first open-source Safety Net implementations for collision-avoidance systems.
DOI: 10.71945/icas2026_0210