Safety assessment for airborne CNN classifier based on conditional Gaussian PAC-Bayes

To address the airworthiness safety challenges caused by inherent uncertainty outputs of machine learning technology in airborne systems, a system safety assessment method based on the generalization theory is proposed for CNN classification under the framework of SAE ARP4761standards. First, based on the PAC-Bayes theory, the training method is improved through conditional gaussian process to optimize the generalization boundand obtain a quantified representation of the uncertainty of the CNN model. Second, an integration method for software uncertaintyand hardware reliability based on generalization bound confidence is proposed to obtain comprehensive failure basic data of CNN components, supporting quantitative safety assessment of the aircraft/system. Finally, taking the airborne GNSS interference signal recognition module based on CNN as a case, the proposed method is shown to be effective in safety assessment, and is also experimentally verified that the generalization boundary based on conditional gaussian process has a tighter computational boundary than that of ordinary PAC-Bayesand Vapnik-Chervonenkis dimensions.