A Survey of Robustness and Safety of 2D and 3D Deep Learning Models Against Adversarial Attacks

Deep learning models used in safety-critical systems like autonomous vehicles and facial recognition are vulnerable to adversarial attacks, meaning they can be fooled by specially crafted inputs that humans would recognize correctly, posing real risks to human safety.

Adversarial threats extend beyond digital manipulation to physical attacks: The survey highlights that adversarial attacks can be "physically realizable," meaning attackers could use real-world modifications (like altered stop signs or objects) to fool AI systems, not just digital perturbations, making these vulnerabilities particularly concerning for deployed systems.

3D models face under-studied security risks: This is the first comprehensive systematic investigation of adversarial attacks on 3D deep learning models, suggesting this emerging area lacks sufficient safety research despite growing real-world applications in autonomous systems and other critical domains.

Current deep learning models lack sufficient robustness: The survey reviews over 170 papers documenting how various adversarial attacks can compromise model reliability, indicating that trustworthiness remains a fundamental challenge before deploying these systems in safety-critical applications.

Building trustworthy AI requires systematic threat modeling: The authors emphasize the need for comprehensive frameworks to understand different types of adversarial threats from multiple perspectives, suggesting that responsible AI deployment demands proactive security analysis rather than reactive fixes after failures occur.