Artificial intelligence (AI)–based perception systems have become a fundamental component of advanced driver assistance and automated driving platforms. These systems rely on deep neural networks (DNNs) to perform safety-critical tasks such as object detection, drivable area segmentation, and lane line estimation. When deployed on embedded hardware, neural networks may experience hardware faults that perturb stored weights without causing immediate system crashes. Such perturbations can silently propagate through the network and degrade perception outputs in subtle yet safety-critical ways. A systematic strategy that considers the interaction between the application and the underlying hardware characteristics remains partially explored and requires structured evaluation frameworks.