A Closer Look at Curriculum Adversarial Training: From an Online Perspective

Authors: Lianghe Shi, Weiwei Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive experiments to support our theory.
Researcher Affiliation Academia School of Computer Science, Wuhan University Institute of Artificial Intelligence, Wuhan University National Engineering Research Center for Multimedia Software, Wuhan University Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is attached to the supplementary material.
Open Datasets Yes Extensive numerical experiments on CIFAR-10 and CIFAR-100 (Krizhevsky, Hinton et al. 2009) datasets verify our theoretical bounds and the explanation we provide regarding curriculum adversarial training.
Dataset Splits No Extensive numerical experiments on CIFAR-10 and CIFAR-100 (Krizhevsky, Hinton et al. 2009) datasets verify our theoretical bounds and the explanation we provide regarding curriculum adversarial training.
Hardware Specification No The paper does not provide specific details on the hardware used, such as GPU/CPU models or memory specifications.
Software Dependencies No For all baselines, we run projected gradient descent (PGD) as our adversary... the models are trained using stochastic gradient descent (SGD)...
Experiment Setup Yes For all baselines, we run projected gradient descent (PGD) as our adversary, with a step size of 0.007. The maximum step of PGD is 20, and the maximum radius of the l -norm bounded perturbation is δ = 0.031. Following Zhang et al. (2020), the models are trained using stochastic gradient descent (SGD) with momentum of 0.9 for 120 epochs. The initial learning rate is 0.1, reduced to 0.01, 0.001, and 0.0005 at epoch 60, 90, and 110, respectively. The batch size is 128.