Standard-Deviation-Inspired Regularization for Improving Adversarial Robustness

Authors: Olukorede Fakorede, Modeste Atsague, Jin Tian

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct an extensive evaluation of the proposed method. To assess its versatility, we test it on various datasets, including CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009), SVHN (Netzer et al., 2011), and Tiny Image Net Deng et al. (2009). We apply simple data augmentations, such as 4-pixel padding with 32 32 random crop and random horizontal flip, to each of the datasets. Additionally, we employ Res Net-18 (He et al., 2016) and Wide Res Net-34-10 (He et al., 2016) as the backbone models.
Researcher Affiliation Academia Olukorede Fakorede EMAIL, EMAIL Department of Computer Science Iowa State University, Ames, Iowa, USA; Modeste Atsague EMAIL Department of Computer Science Iowa State University, Ames, Iowa,USA; Jin Tian EMAIL Mohamed bin Zayed University of Artificial Intelligence Abu Dhabi, United Arab Emirates
Pseudocode Yes Algorithm 1 AT-SDI Algorithm. Input: a neural network model with the parameters θ, step size κ, T PGD steps, a training dataset D of size n, |C| is the number of classes, and hyperparameter β. Output: a robust model with parameters θ; Algorithm 2 SDI-PGD Algorithm. Input: a neural network model with the parameters θ, step size κ, natural examples xi in a labelled dataset D of size n and |C| is the number of classes. Output: Adversarial examples x i
Open Source Code No The paper does not provide an explicit statement about open-sourcing its code or a link to a code repository.
Open Datasets Yes To assess its versatility, we test it on various datasets, including CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009), SVHN (Netzer et al., 2011), and Tiny Image Net Deng et al. (2009).
Dataset Splits No We train the backbone networks using mini-batch gradient descent for 110 epochs, with a momentum of 0.9 and a batch size of 128. For training CIFAR-10, we used a weight decay of 5e-4, and for CIFAR-100, SVHN, and Tiny Image Net, we used a weight decay of 3.5e-3. The initial learning rate was set to 0.1 (0.01 for CIFAR-100, SVHN, and Tiny Image Net), and it was divided by 10 at the 75th epoch and then again at the 90th epoch. The hyperparameters are tuned using a validation set.
Hardware Specification Yes We conducted all experiments using a single core of an AMD EPYC 7513 processor, an Nvidia A100 SXM4 80 GB GPU, and 128 GB of RAM.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes We train the backbone networks using mini-batch gradient descent for 110 epochs, with a momentum of 0.9 and a batch size of 128. For training CIFAR-10, we used a weight decay of 5e-4, and for CIFAR-100, SVHN, and Tiny Image Net, we used a weight decay of 3.5e-3. The initial learning rate was set to 0.1 (0.01 for CIFAR-100, SVHN, and Tiny Image Net), and it was divided by 10 at the 75th epoch and then again at the 90th epoch.