Adversarial Examples in Environment Perception for Automated Driving (Review)

Deep neural networks powering self-driving cars can be fooled by tiny, invisible changes to images — adversarial perturbations imperceptible to human eyes can cause AI systems to misidentify objects, pedestrians, or road signs, creating serious safety risks for autonomous vehicles.

This vulnerability represents a critical gap between human and machine perception — while humans can easily recognize a stop sign or pedestrian despite minor visual distortions, the neural networks controlling self-driving cars may fail under similar conditions, highlighting why current AI systems cannot yet be fully trusted for safety-critical applications.

Both attack and defense methods are actively being researched — the paper reviews a decade of work on both how adversarial examples can be created and how AI systems can be made more robust against them, indicating this is an ongoing challenge rather than a solved problem.

Adversarial robustness is essential for trustworthy autonomous driving — addressing these vulnerabilities is not optional but fundamental to developing AI systems that can be safely deployed on public roads, making this research critical for responsible AI development in transportation.

The field is still developing solutions — by surveying the research landscape, this review suggests that comprehensive defenses against adversarial attacks in autonomous driving remain an active area of investigation without definitive solutions yet established.