Neural networks (NNs) have shown remarkable performance of perception in their application in autonomous vehicles (AVs). However, NNs are intrinsically vulnerable to perturbations, such as occurrences outside of the training sets, scene noise, instrument noise, image translation, and rotation, or small changes intentionally added to the original image (called adversarial perturbations). Incorrect conclusions from the perception systems (e.g., missing objects, wrong classification, and traffic sign misdetection or misreading) have been a major cause of disengagement incidents in AVs. In order to explore the dynamic nature of hazardous events in AVs, we develop a range of methods to analyze AV safety and security. This work is part of the project and is devoted to analyzing the influence of robustness in the NN-based perception system by using fault tree analysis (FTA). We extend the traditional FTA to represent combinations of failure causes in the multi-dimensional space, i.e., two variables that influence whether the image is classified correctly. The extended FTA is demonstrated on the traffic sign recognition module of AV theoretically and in practice.