AI Framework Simulates Traffic Sign Decay to Expose Autonomous Driving Flaws

Researchers from Seoul National University of Science and Technology (SEOULTECH) and Kyung Hee University have developed a new framework designed to stress-test the vision systems used in autonomous vehicles. The system, called Adversarial Wear and Tear (AdvWT), simulates how natural environmental decay can trick AI into misidentifying critical road signage.

The AdvWT framework utilizes a generative image-to-image translation model based on StarGAN-v2. By learning a latent "damage style," the system can reproduce realistic physical deterioration on traffic signs while ensuring the original identity of the sign remains intact. In testing, the framework achieved near-perfect attack success rates when targeting lightweight convolutional neural networks (CNNs), specifically ResNet-18 and MobileNet, proving that simulated decay can effectively mislead AI vision.

The Vulnerability of Decay

Deep neural network (DNN)-based vision systems are now central to the deployment of autonomous driving. While the industry has extensively studied temporary optical attacks—such as lasers or strategic stickers—permanent environmental degradation remains a persistent threat. Unlike these fleeting interventions, natural wear and tear is a constant factor in real-world infrastructure. Dr. Hong Joo Lee noted that this type of deterioration persists until the physical object is eventually repaired or replaced, creating a long-term vulnerability in safety-critical AI deployments.

Implications for AI Reliability

This research exposes a significant gap in current AI reliability: models that demonstrate high accuracy on clean, curated datasets often fail when confronted with the unpredictability of the physical world. When an autonomous vehicle misinterprets a faded or rusted stop sign, the result can be catastrophic. By using AdvWT to intentionally identify these failure points, developers can move beyond idealized data and understand exactly how environmental degradation triggers system errors.

Strengthening Safety-Critical Systems

Beyond exposing flaws, the AdvWT framework provides a pathway toward safer AI. The realistic simulations of damaged signs can be integrated into training datasets, allowing AI models to learn the patterns of decay and become more resilient to physical-world damage. This approach to "adversarial training" is expected to reduce the risk of accidents in autonomous transport, as well as other safety-critical sectors such as healthcare and finance, where AI must operate reliably in imperfect environments.