Reproduce “Corrupting Attention: Evasion-Based Adversarial Attacks on Encoder Attention in Detection Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ridma Jayasundara, Shaheer Mohamed, Tharindu Fernando, Harshala Gammulle, Basura Fernando, Sanka Rasnayake, A V Subramanyam, Sridha Sridharan and 1 more
The authors introduce an attack that disrupts object detectors by corrupting encoder attention rather than final detections. It optimizes bounded image perturbations against structured targets for attention rows linked to predicted objects. On COCO, DETR-R50 mAP falls from 42.1 clean to 0.97, versus AFOG’s 4.1 under the same budget and 10 iterations; this shows greater damage in that setting, not sole causation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.