Synthesizing Robust Physical Camouflage for Universal 3D Evasion Attacks

Abstract:

The rise of deep learning in safety-critical domains like autonomous vehicles and surveillance systems has underscored the urgency of protecting these technologies against adversarial attacks. Physical adversarial attacks, which involve tangible modifications to an object’s appearance, present a practical and significant threat since they can induce misinterpretation of computer vision models. This dissertation demonstrates this critical issue by developing methodologies for generating physical adversarial camouflage and providing comprehensive preliminaries to deepen the understanding of the field.

The research is anchored in two interconnected goals. The first is developing a novel Neural Renderer, addressing the challenge of non-differentiability in legacy 3D simulators. This innovation is crucial for optimizing adversarial textures in a white-box attack framework, enabling the use of accurate and photo-realistic scenes in the optimization pipeline.

The second goal involves formulating two sophisticated frameworks for generating adversarial textures. These frameworks leverage the previously developed Neural Renderer and introduce new loss functions and modules to enhance attack effectiveness. Their aim is to produce universally effective textures across various conditions and models, demonstrating both the versatility and potential threats of such adversarial techniques.

Collectively, this dissertation makes a significant contribution to the field of adversarial machine learning. It not only presents innovative methods for physical adversarial camouflage but also highlights the vulnerabilities in AI systems, emphasizing the urgent need for enhanced AI security in critical applications.