AdvFMEA: Adversarial-Aware Failure Mode and Effects Analysis for Safety-Critical Monocular Depth Estimation in Autonomous Vehicles

Document Type

Article

Publication Title

IEEE Access

Abstract

The reliability of monocular depth estimation is paramount for safe navigation in autonomous vehicles. However, these systems face critical safety vulnerabilities from physical adversarial attacks, such as road marking manipulations and object-blending perturbations. This work introduces an adversarial-aware Failure Mode and Effects Analysis (AdvFMEA) framework that quantifies these risks through an Adversarially Aware Risk Priority Number (AA-RPN) integrating factors such as attack persistence, depth-error severity, and adversarial impact. Benchmarking seven state-of-the-art models across CARLA simulations and the real-world KITTI dataset reveals that transformer architectures suffer from systemic depth spoofing. At the same time, lightweight CNNs exhibit higher safety-critical error rates. Our fail-operational approach includes an attack-agnostic FMEA methodology compliant with ISO 21448 (SOTIF), empirical validation of sim-to-real threat correlation, and deployable safety thresholds for resource-constrained systems.

First Page

203230

Last Page

203252

DOI

10.1109/ACCESS.2025.3638855

Publication Date

1-1-2025

This document is currently not available here.

Share

COinS