Sharp Statistical Guaratees for Adversarially Robust Gaussian Classification

Authors: Chen Dan, Yuting Wei, Pradeep Ravikumar

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Figure 1. A simple simulation on the performance of Algorithm 1 and the algorithm proposed in (Schmidt et al., 2018) is shown here with different values of Adv SNR r. Here we consider a 50-dimensional example under ℓ∞ adversary with ε = 0.1. The covariance matrix is fixed to be Σ = I, and the mean parameter µ is set as µ = (r + ε, ε, ε, . . . , ε) for r ∈ {0.5, 1.0, 2.0}. We evaluate the excess risk RB,ε µ,Σ(f ˆw) − RB,ε µ,Σ returned by the two algorithms using n i.i.d. training data pairs, where n ∈ {100, 200, 400, 800, 1600, 3200, 6400, 12800}. For each combination of (n, r), the averaged excess risk over 10 random repetitions is reported respectively.
Researcher Affiliation Academia 1Carnegie Mellon University, Pittsburgh, Pennsylvania, USA.
Pseudocode Yes Algorithm 1 A plug-in estimator of w0
Open Source Code No The paper does not include any explicit statements about releasing code or links to a code repository for the methodology described.
Open Datasets No The paper describes using a "Gaussian mixture model proposed by Schmidt et al. (2018)" for data generation, and refers to "n i.i.d. training data pairs" in a simulation (Figure 1). However, it does not provide concrete access information (e.g., a specific link, DOI, repository name, or formal citation with authors/year for a publicly available dataset) for the data used in the experiments. The data appears to be simulated or generated rather than a standard public dataset.
Dataset Splits No The paper mentions "n i.i.d. training data pairs" and evaluation "at the testing stage" but does not specify training, validation, or test dataset splits (e.g., percentages or counts) or reference standard predefined splits.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU/CPU models or cloud instances.
Software Dependencies No The paper does not specify any software dependencies (e.g., libraries, frameworks, or solvers) with version numbers that would be needed to reproduce the experiments.
Experiment Setup Yes Here we consider a 50-dimensional example under ℓ∞ adversary with ε = 0.1. The covariance matrix is fixed to be Σ = I, and the mean parameter µ is set as µ = (r + ε, ε, ε, . . . , ε) for r ∈ {0.5, 1.0, 2.0}. We evaluate the excess risk RB,ε µ,Σ(f ˆw) − RB,ε µ,Σ returned by the two algorithms using n i.i.d. training data pairs, where n ∈ {100, 200, 400, 800, 1600, 3200, 6400, 12800}. For each combination of (n, r), the averaged excess risk over 10 random repetitions is reported respectively.