SegPAR for Efficient Sparse Attacks on Semantic Segmentation
August 13, 2026
SegPAR introduces a class-centric exploration paradigm and a discrepancy reward to improve sparse decision-based black-box attacks. This approach addresses the query inefficiency found in existing image-centric pixel accumulation methods for semantic segmentation.
HOW THIS AFFECTS YOU
●
researcherYou can utilize this class-centric framework to conduct more efficient black-box vulnerability testing on segmentation models.