Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Adversarial Robustness of Nonparametric Regression
Authors: Parsa Moradi, Hanzaleh Nodehi, Mohammad Ali Maddah-Ali
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present numerical experiments to validate the theoretical results. All experiments are conducted on a single CPU-only machine. The smoothing spline estimator is implemented using the SciPy package [58]. We consider two regression functions: (i) f(x) = x sin(x) over the domain Ω= [ 10, 10] with M = 100, and (ii) a three-layer MLP network with weights initialized in [ 1, 1] and M = 500. The noise vector ε is drawn independently from a Gaussian distribution with zero mean and variance σ2 = 1. Figures 3 and 5 illustrate the behavior of R2(f, ˆf a SS) and R (f, ˆf a SS) under uniform and Gaussian designs, respectively, for two corruption levels, q = n0.3 and q = n0.6. Similarly, for the MLP network, Figures 4 and 6 present the corresponding results. |
| Researcher Affiliation | Academia | Parsa Moradi University of Minnesota EMAIL Hanzaleh Akbarinodehi University of Minnesota EMAIL Mohammad Ali Maddah-Ali University of Minnesota EMAIL |
| Pseudocode | No | The paper describes the methodology and algorithms in prose and mathematical formulations within the main text and appendices. No explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | No | We provided references to all packages that we used in the paper. Regarding the code, we are happy to share it later if required. |
| Open Datasets | No | The paper uses synthetic data generated based on specific regression functions (f(x) = x sin(x) or a three-layer MLP network) with added Gaussian noise. The design points are either uniform or Gaussian distributions over the domain. No external publicly available datasets are utilized or referenced. |
| Dataset Splits | No | The paper uses synthetic data generated based on specific regression functions and design points (uniform and Gaussian). It does not describe any train/test/validation splits typically used for pre-existing datasets. The experiments involve varying sample sizes 'n' and corruption levels 'q'. |
| Hardware Specification | No | All experiments are conducted on a single CPU-only machine. |
| Software Dependencies | No | The smoothing spline estimator is implemented using the SciPy package [58]. |
| Experiment Setup | Yes | We consider two regression functions: (i) f(x) = x sin(x) over the domain Ω= [ 10, 10] with M = 100, and (ii) a three-layer MLP network with weights initialized in [ 1, 1] and M = 500. The noise vector ε is drawn independently from a Gaussian distribution with zero mean and variance σ2 = 1. To evaluate the adversarial robustness of the cubic smoothing spline estimator, we consider three distinct attack strategies: Random Corruption Attack, Greedy Corruption Attack, Concentrated Corruption Attack. For each attack strategy, we evaluate both R2(f, ˆf a SS) and R (f, ˆf a SS) across a range of sample sizes n, and examine how these metrics scale with n under varying levels of adversarial corruption. Additionally, for each experiment, we consider two settings for the design points: uniform and Gaussian. For two corruption levels, q = n0.3 and q = n0.6. |