Reproduce “TOG: Targeted Adversarial Objectness Gradient Attacks on Real-time Object Detection Systems” 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
Ka-Ho Chow, Ling Liu, Mehmet Emre Gursoy, Stacey Truex, Wenqi Wei, Yanzhao Wu
The paper introduces TOG, a framework of targeted adversarial attacks designed to exploit vulnerabilities in real-time object detection systems like YOLO and SSD. The research demonstrates that imperceptible perturbations can cause objects to vanish, false positives to appear, or objects to be mislabeled, significantly degrading detector performance with mAP reductions to near zero in some cases.
Key insights
- Object detection systems possess a larger and more complex attack surface compared to image classifiers due to their multi-output nature (object existence, localization, and classification).
- Adversarial examples crafted for one object detector can successfully transfer to other detectors, including those of different architectures, enabling black-box attacks.
- Universal adversarial perturbations can be pre-computed offline and applied efficiently online, posing a significant threat to resource-constrained edge deployments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.