STONE: Pioneering the One-to-N Backdoor Threat in 3D Point Cloud
This work addresses a critical security threat for safety-sensitive domains like autonomous driving and robotics by establishing a foundational benchmark for multi-target backdoor attacks in 3D vision.
The paper tackles the problem of one-to-N backdoor attacks in 3D point clouds, which were previously unexplored, by introducing the STONE framework with a spherical trigger, achieving up to 100% attack success rate without compromising clean-data accuracy.
Backdoor attacks pose a critical threat to deep learning, especially in safety-sensitive 3D domains such as autonomous driving and robotics. Despite their potency, existing attacks on 3D point clouds are limited to a static one-to-one paradigm, leaving the more flexible one-to-N backdoor threat largely unexplored and without a theoretical or practical foundation. We address this by introducing STONE (Spherical Trigger One-to-N Backdoor Enabling), the first framework that instantiates this threat through a configurable spherical trigger. Its parameterizable spatial properties create a dynamic key space, enabling a single trigger to control multiple output labels. Theoretically, we ground STONE through Neural Tangent Kernel (NTK) analysis, providing the first formal basis for one-to-N mappings in 3D models. Empirically, extensive evaluations show high attack success rate (up to 100\%) with no loss in clean-data accuracy. This work establishes a foundational benchmark for multi-target threats in 3D vision, crucial for securing future intelligent systems.