Consistent Adversarially Robust Linear Classification: Non-Parametric Setting
Authors: Elvis Dohmatob
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present some empirical verification of our theoretical results. All experiments were run on a single modern CPU laptop. |
| Researcher Affiliation | Industry | Elvis Dohmatob 1 1Meta FAIR. Correspondence to: Elvis Dohmatob <dohmatob@meta.com>. |
| Pseudocode | No | The paper describes the algorithm in text but does not provide a formal pseudocode block or a clearly labeled 'Algorithm' section with structured steps. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code for the described methodology or a link to a code repository. |
| Open Datasets | No | The paper describes synthetic data generation processes (e.g., 'Consider the distribution P(y = 1) = 1/2, x | y N(yµ, Σ)' and 'z | y N(yµ, Id), x = max(z, 0)') rather than referencing or providing concrete access information for a publicly available or open dataset. |
| Dataset Splits | No | The paper mentions generating training data ('Dn = {(x1, y1), . . . , (xn, yn)} of n iid samples') and evaluating adversarial risk, but it does not specify explicit train/validation/test splits, percentages, or a cross-validation methodology. |
| Hardware Specification | No | All experiments were run on a single modern CPU laptop. |
| Software Dependencies | No | We use trust-region-based methods (Absil et al., 2007) implemented in the Manopt library (Boumal et al., 2014). |
| Experiment Setup | Yes | We set the input-dimension to d = 20 for this experiment. For each value of sample size n {100, 200, . . . , 1000, 2000, 3000, . . . , 10000}, we generate a dataset Dn = {(x1, y1), . . . , (xn, yn)} of n iid samples, and then compute the estimator bfn,ϵ,hn described in Section 3.3, where hn is the bandwidth parameter, taken as hn = p (d/n) log n, in accordance with the choice in Corollary 4.1. We consider Euclidean-norm attacks of strength ϵ ranging in {0.1, 0.2, . . . , 0.9, 1}. |