ScalePerson: Towards Good Practices in Evaluating Physical Adversarial Attacks on Person Detection
Keywords: Physical Adversarial Attack, Person Detection, Dataset
TL;DR: Our paper introduces ScalePerson, a new dataset and benchmark for evaluating physical adversarial attacks in person detection, providing standardized metrics and comprehensive analyses across multiple attack methods and detectors.
Abstract: Person detection is widely used in safety-critical tasks but is known to be vulnerable to physical adversarial attacks. Numerous pioneering attack methods have been proposed, each claiming superior performance and exposing potential security risks. However, assessing actual progress in this field is challenging due to two common limitations in existing evaluations. First, inconsistent experimental setups and ambiguous evaluation metrics hinder fair comparisons. Second, the absence of a dedicated dataset for this task has led to evaluations on datasets originally designed for object detection, which, while informative, are inadequate. To address these limitations, we present a comprehensive benchmark and introduce ScalePerson, the first dataset specifically designed for evaluating physical adversarial attacks in person detection. This dataset incorporates critical factors for this task, such as person scale, orientation, number of individuals, and capture devices. Our benchmark includes standardized evaluation metrics and a modular codebase to enhance reproducibility and transparency. Leveraging this benchmark, we conduct an extensive evaluation of 11 state-of-the-art attacks against 7 mainstream detectors across 3 datasets, totaling 231 experiments. We present detailed analyses from multiple perspectives, examining the impact of various factors on the efficacy of physical adversarial attacks in person detection. The source code and dataset will be made publicly available upon acceptance of this paper.
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5604
Loading