Formal Verification of Neural Networks: A "Step Zero" Approach for Vehicle Detection

This paper delves into the verification of Convolutional Neural Networks for the crucial task of identifying vehicles in automotive images. Given the complexity and verifiability challenges of traditional object detection models, we propose a "step zero" approach, focusing on certifying the robustness of classification models for vehicle recognition. Our research paves the way for utilising these certified models as a potential safety net in future applications. While not yet empirically tested alongside object detection models, this approach offers promising prospects for reducing the risk of false negatives, contributing to the development of dependable AI systems in the automotive domain.