Skip to content

Patch Attacks Against Object Detectors Using Out-of-Distribution Smoothing Defenses

Defending Object Detectors against Patch Attacks with Out-of-Distribution Smoothing

arXiv· arXiv:2205.08989· Ryan Feng, Neal Mangaokar, Jihye Choi, Somesh Jha, Atul PrakashUniversity of Michigan; University of WisconsinIndexed by this site · 23 days ago· Originally publishedDefenses

Translation in progress — untranslated text is omitted.

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

AI summaryThe paper proposes the OODSmoother framework to uniformly characterize defense methods aimed at removing adversarial patches, and based on it designs an adaptive attack that can break existing patch defenses and a SemPrior defense that leverages semantic priors. The authors state that SemPrior alone brings up to 40% improvement, and when combined with existing defenses brings up to 60% improvement.

Questions
Object detectors face physically realizable patch attacks; existing defenses mostly rely on training a patch segmenter to detect and remove the patch, and the authors argue that such defenses lack a unified characterization and that ML patch detectors struggle to learn semantic priors.
Methods
The authors propose the OODSmoother framework, which unifies such defenses as an OOD oracle plus an OOD smoother; based on this, they design adaptive attacks that use two losses to attack the patch segmenter and the downstream detector respectively, and propose a SemPrior defense based on color histograms and HSV statistics.
Experiments and results
Under 100×100 patches, the adaptive attack reduces SAC's mAP@IoU=0.5 on COCO from 53.6% to 6.2%, and PatchZero's on VOC from 68.9% to 7.0%; the authors report that SemPrior alone improves by up to 40%, and combined with SAC by up to 60%.
Limitations and future work
The authors note that SemPrior may mistakenly affect natural objects using rare colors, and that monochromatic concentrated images are prone to false positives; the experiments cover only gradient attacks and do not consider diffusion-generated attacks; the downstream detector tested is Faster RCNN (R50), with 1000 images evaluated per setting.

Related papers

Show 3 more papers
Read originalarxiv.org