Formal Verification of Deep Neural Networks for Image Classification in Industrial Environments

Abstract

Deep Learning methods, particularly Convolutional Neural Networks (CNNs), have become widely adopted in industrial computer vision applications due to their strong image classification capabilities. Industrial autonomous systems, such as autonomous mobile robots (AMRs), rely on accurate object perception to ensure safe operation and reliable decision-making in complex environments. However, despite their high performance, CNN-based models remain vulnerable to adversarial perturbations and sensor-related degradations, which may lead to incorrect classifications and compromise system reliability in safety-critical applications.This work presents the robustness of CNN-based image classification models for industrial applications by analyzing their performance under increasing perturbation levels and identifying the limits at which misclassification occurs. An industrial object dataset containing seven classes was constructed using an object detection and extraction pipeline, resulting in object-centered RGB and depth image samples. To perform this analysis, two CNN-based architectures were developed and evaluated: a baseline RGB model and a multimodal RGB-D model incorporating depth information from multiple sources.Robustness was assessed through two complementary approaches. Formal verification was performed using the alphabeta-CROWN framework under an L∞-bounded threat model, providing mathematical guarantees on model predictions under bounded perturbations. In addition, corruption-based experiments following the ImageNet-C protocol were conducted to evaluate model behavior under common image degradations, with a depth-suitable corruption subset introduced for metric depth modalities.