Adv-Attribute: Inconspicuous and Transferable Adversarial Attack on Face Recognition

Authors: Shuai Jia, Bangjie Yin, Taiping Yao, Shouhong Ding, Chunhua Shen, Xiaokang Yang, Chao Ma

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the FFHQ and Celeb A-HQ datasets show that the proposed Adv-Attribute method achieves the state-of-the-art attacking success rates while maintaining better visual effects against recent attack methods.
Researcher Affiliation Collaboration Shuai Jia1 Bangjie Yin2 Taiping Yao2 Shouhong Ding2 Chunhua Shen3 Xiaokang Yang1 Chao Ma1 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Youtu Lab, Tencent 3 Zhejiang University
Pseudocode Yes Algorithm 1: The proposed Adv-Attribute. Input: Source image xo Xo; Victim image xt Xt; Targeted FR model f( ); Pre-trained Style GAN generator Gs; Adversarial noise generator Gi a, i {1, ..., N}. Output: Adversarial examples xadv;
Open Source Code No The paper does not provide an explicit statement about the availability of its source code or a link to a code repository.
Open Datasets Yes In our work, we choose two public facial datasets for evaluation: 1) FFHQ [29] is a high-quality human face dataset, which consists of more than 70,000 high-quality human face images with variations of age and ethnicity. 2) Celeb A-HQ [33] is constructed as a higher-quality version of the Celeb A dataset [34].
Dataset Splits No The paper specifies the selection of 100 source faces and 10 target faces to construct 1000 source-target pairs for the impersonation attack. However, it does not explicitly provide the training, validation, and test splits for the dataset used to train their Adv-Attribute framework or the attribute noise generators.
Hardware Specification Yes The overall framework is implemented in Py Torch on one NVIDIA Tesla P40 GPU.
Software Dependencies No The paper mentions 'implemented in Py Torch' and 'trained using Adam optimizer [36]' but does not provide specific version numbers for PyTorch or any other software libraries or dependencies.
Experiment Setup Yes The attribute noise generators are trained using Adam optimizer [36] with an initial learning rate of 0.0001.