Human-Imperceptible, Machine-Recognizable Images

Authors: Fusheng Hao, Fengxiang He, Yikai Wang, Fuxiang Wu, Jing Zhang, Dacheng Tao, Jun Cheng

IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on Image Net and COCO show that the proposed paradigm achieves comparable accuracy with the competitive methods.
Researcher Affiliation Academia Fusheng Hao1,2 , Fengxiang He3 , Yikai Wang4 , Fuxiang Wu1,2 , Jing Zhang5 Dacheng Tao6 , Jun Cheng1,2 1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences 2The Chinese University of Hong Kong 3University of Edinburgh 4Beijing Normal University 5The University of Sydney 6Nanyang Technological University
Pseudocode Yes The pseudocodes of RS and MI in a Py Torch-like style are shown in the appendix.
Open Source Code Yes Co-corresponding authors Code: https://github.com/Fusheng Hao/Privacy Preserving ML
Open Datasets Yes For the image classification task, we benchmark the proposed PEVi T on Image Net-1K [Deng et al., 2009], which contains 1.28M training images and 50K validation images. For the object detection task, we benchmark the proposed PEYOLOS on COCO [Lin et al., 2014], which contains 118K training, 5K validation and 20K test images.
Dataset Splits Yes For the image classification task, we benchmark the proposed PEVi T on Image Net-1K [Deng et al., 2009], which contains 1.28M training images and 50K validation images. For the object detection task, we benchmark the proposed PEYOLOS on COCO [Lin et al., 2014], which contains 118K training, 5K validation and 20K test images.
Hardware Specification Yes The throughput is measured as the number of images processed per second on a V100 GPU. FPS is measured with batch size 1 on a single 1080Ti GPU. We adopt the default hyper-parameters of the Dei T training scheme [Touvron et al., 2020] except setting the batch size to 192 per GPU, where 8 NVIDIA A100 GPUs are used for training.
Software Dependencies No The paper mentions software by name (Timm library, publicly released code in [Fang et al., 2021]) but does not provide specific version numbers for these, nor does it list multiple key software components with their versions.
Experiment Setup Yes We adopt the default hyper-parameters of the Dei T training scheme [Touvron et al., 2020] except setting the batch size to 192 per GPU, where 8 NVIDIA A100 GPUs are used for training.