Researchers Develop AdvWT to Test AI Traffic Sign Recognition

Seoul National University of Science and Technology announced in a press release that researchers developed Adversarial Wear and Tear, or AdvWT, a framework for testing AI vision systems with realistic traffic sign damage. The system adds fading, cracks, and corrosion to traffic signs to check whether recognition models misread them.

The research team included Associate Professor Seong Tae Kim from Kyung Hee University and Assistant Professor Hong Joo Lee from Seoul National University of Science and Technology. The findings were made available online on February 3, 2026, and published on May 12, 2026, in IEEE Transactions on Dependable and Secure Computing.

AdvWT uses a generative image to image translation model based on StarGAN-v2 to learn damage patterns while keeping the sign identity intact. The team tested the framework on two traffic sign datasets and eight recognition architectures. It achieved near perfect attack success rates on lightweight CNNs such as ResNet-18 and MobileNet, and also affected transformer models.

The researchers also printed clean and adversarial speed limit signs, then photographed them at different distances, angles, and indoor and outdoor settings. The resulting images still misled the classifier. Training models with signs generated by AdvWT improved their ability to handle naturally damaged traffic signs.