SegPAR: New Sparse Attack Method for Semantic Segmentation
A new research paper on arXiv introduces SegPAR, a decision-based black-box sparse attack framework for semantic segmentation. The paper, identified as arXiv:2608.11285, addresses the limited attention sparse decision-based black-box threats have received in semantic segmentation. The authors adapt representative decision-based black-box sparse attacks from classification to serve as baselines, establishing a benchmark for this underexplored setting. They demonstrate that one existing method suffers from severe query inefficiency due to image-centric pixel accumulation, which exhausts query budgets. SegPAR shifts to a class-centric exploration paradigm and introduces a novel discrepancy reward to eliminate misleading feedback. Experiments show SegPAR significantly outperforms black-box baselines in sparsity efficiency.
Key facts
- Paper arXiv:2608.11285 announces SegPAR, a decision-based black-box sparse attack for semantic segmentation.
- The paper establishes a benchmark by adapting classification-based sparse attacks to semantic segmentation.
- Existing methods suffer from query inefficiency due to image-centric pixel accumulation.
- SegPAR uses a class-centric exploration paradigm.
- A novel discrepancy reward is introduced to improve feedback.
- Experiments show SegPAR outperforms black-box baselines in sparsity efficiency.
- The paper is published on arXiv with announcement type 'cross'.
- The research addresses the gap in sparse decision-based black-box threats in semantic segmentation.
Entities
Institutions
- arXiv