Class-Disentanglement and Applications in Adversarial Detection and Defense

Authors: Kaiwen Yang, Tianyi Zhou, Yonggang Zhang, Xinmei Tian, Dacheng Tao

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, this simple approach substantially improves the detection and defense against different types of adversarial attacks.
Researcher Affiliation Collaboration University of Science and Technology of China1; University of Washington, Seattle2 University of Maryland, College Park3; JD Explore Academy4
Pseudocode No The paper includes an architecture diagram (Figure 1) and mathematical equations for the objective function and attacks, but no structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available: https://github.com/kai-wen-yang/CD-VAE.
Open Datasets Yes all the experiments are conducted on two datasets, CIFAR-10 [26] and restricted Image Net [44].
Dataset Splits No The paper provides training and test set sizes ('50,000 training images and 10000 test images' for CIFAR-10, and '257,748 training images and 10,150 test images' for restricted Image Net) but does not specify details for a separate validation split.
Hardware Specification No The paper mentions 'GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC' but does not provide specific GPU models, CPU details, or other hardware specifications.
Software Dependencies No The paper mentions 'Adam W' as the optimizer but does not specify versions for software libraries, frameworks, or other dependencies like Python, PyTorch, or TensorFlow.
Experiment Setup Yes For CIFAR-10, we use a VAE G( ) with a few convolutional layers [25] and Wide Res Net-28-10 [48] as the image classifier D( ). For restricted Image Net, due to the high resolution (299 299), we use VQ-VAE [38] as G( ) and Res Net-50 [19] as D( ). We set β = 0.2. During training, we use Adam W [12] with a weight decay of 1e-6 as the optimizer to minimize Eq. (1) in an end-to-end manner for 300(60) epochs on CIFAR-10(restricted Image Net).