Researchers have introduced SegPAR, a new framework designed to improve sparse decision-based black-box attacks for semantic segmentation. Existing methods struggle with query inefficiency due to their image-centric approach, which quickly depletes query budgets. SegPAR addresses this by adopting a class-centric exploration paradigm and incorporating a novel discrepancy reward to mitigate misleading feedback during pixel accumulation. Experiments indicate that SegPAR surpasses current black-box baselines in sparsity efficiency and MIoU reduction, while remaining competitive with white-box attacks. AI
IMPACT This research could lead to more robust semantic segmentation models by improving adversarial attack techniques.
RANK_REASON The cluster contains an academic paper detailing a new method for adversarial attacks in computer vision. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
