Adversary ML Resilience in Autonomous Driving Through Human Centered Perception Mechanisms

Physical attacks on road signs pose real safety risks to autonomous vehicles: The paper demonstrates that simple adversarial attacks—tape, graffiti, and changes in lighting—can cause AI systems to misclassify critical road signs like stop signs and speed limits, highlighting a concrete vulnerability in current autonomous driving technology.

Standard AI models struggle to handle real-world variations: The research shows that machine learning models often either memorize training data without adapting to new situations (overfitting) or fail to learn enough to classify signs accurately (underfitting), meaning AVs trained in controlled conditions may fail in unpredictable real-world environments.

Transfer learning improves robustness across different scenarios: A key finding is that models trained on simple geometric shapes (circles, squares, triangles) significantly improved their ability to correctly identify road signs, suggesting that teaching AI systems broader visual concepts helps them handle unexpected variations and attacks better.

Adversarial training is an effective defense strategy: By deliberately exposing AI models to adversarial attacks during training, the systems became more resilient to these same attacks during real-world deployment—a practical approach to building safer autonomous vehicles.

Human-centered design should guide future AV safety research: The paper's emphasis on human-centered perception mechanisms suggests that developing autonomous vehicles requires not just technical fixes, but approaches that account for how humans understand and interact with road environments.