Revisiting Adversarial Patches for Designing Camera-Agnostic Attacks against Person Detection

Authors: Hui Wei, Zhixiang Wang, Kewei Zhang, Jiaqi Hou, Yuanwei Liu, Hao Tang, Zheng Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our proposed Camera-Agnostic Patch (CAP) attack effectively conceals persons from detectors across various imaging hardware, including two distinct cameras and four smartphones.
Researcher Affiliation Academia 1National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University 2The University of Tokyo 3School of Computer Science, Peking University
Pseudocode Yes Algorithm 1 The proposed adversarial optimization ( Attacker and Defender)
Open Source Code No The source code will be made available upon acceptance of the paper.
Open Datasets Yes We use the INRIAPERSON dataset [4, 30] to evaluate digital-space attacks. ... YOLOv5 [17] model pre-trained on the COCO dataset [20]
Dataset Splits No The paper mentions 613 training images and 288 test images for the INRIAPerson dataset but does not explicitly specify a validation set split.
Hardware Specification Yes Our implementation utilizes PyTorch on a Linux server equipped with dual NVIDIA GeForce RTX 3090 GPUs.
Software Dependencies No Our implementation utilizes PyTorch. However, no specific version number for PyTorch or any other software dependency is provided.
Experiment Setup Yes The adversarial patches are configured with dimensions of 300 × 300, and we employ a YOLOv5 [17] model pre-trained on the COCO dataset [20] and subsequently fine-tuned on INRIAPerson [30] as our victim detector. The detector processes input images at a resolution of 640 × 640, and adversarial training proceeds for 100 epochs.