High-dimensional (Group) Adversarial Training in Linear Regression

Authors: Yiling Xie, Xiaoming Huo

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we will run numerical experiments to observe the empirical performances of (group) adversarial training in high-dimensional linear regression.
Researcher Affiliation Academia Yiling Xie School of Industrial and Systems Engineering Georgia Institute of Technology Atlanta, Georgia, USA yxie350@gatech.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our numerical experiments are based on synthetic data. The code has been submitted in the Supplementary Materials.
Open Datasets No We consider the following models to generate synthetic data: The response variable Y is generated by the Gaussian linear model, as stated in Assumption 1.2.
Dataset Splits No The paper mentions sample sizes for synthetic data generation but does not specify explicit train/validation/test dataset splits.
Hardware Specification No The paper does not provide any specific details about the hardware used for running the experiments.
Software Dependencies No In terms of computation, we apply the dual formulations, i.e., problem (2) and problem (8), and solve these convex optimization problems using the CVXPY toolbox [8]. (No version number for CVXPY is specified).
Experiment Setup Yes The standard deviation of the error ϵ is chosen as 0.1. [...] the perturbation magnitude is chosen in the order of 1/ n in the adversarial training; the ratio of the perturbation magnitude and the perturbation weight is chosen in the order of 1/ n in the group adversarial training. For the constant, we selected 1 for simplicity and experimental convenience. For the group adversarial training, we divide the parameter equally into 125 groups of size 4 for Model 1 and 200 groups of size 3 for Model 2. The sample sizes are chosen {50, 100, 150, 200, 250, 300, 350, 400} for Model 1 and {50, 100, 150, 200, 250, 350, 450, 550} for Model 2.