TwinIR: Coordinated Invisible Dual-Point Attacks on Online HD Map Construction

A recent study published on arXiv (ID: 2608.04453) presents TwinIR, a method for executing physical attacks on the creation of online HD maps used in autonomous vehicles. The researchers reveal a phenomenon termed the 'cross-boundary compensation effect,' which hinders current physical attacks; after altering one boundary, other visible boundaries may still provide geometric information that helps the model restore the original road layout. TwinIR aims to optimize both the effectiveness of the attack and the sparsity of points, determining the fewest attack points necessary to diminish these compensating cues. To minimize visibility, it simulates camera responses to near-infrared light and strategically places attack points to create interference that is noticeable to cameras but not in the visible spectrum. Experiments conducted in the study are referenced but not fully detailed. This research underscores the security risks associated with autonomous driving systems and their map construction processes.