Adversarial Surrogate Risk Bounds for Binary Classification
Authors: Natalie Frank
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | This paper provides surrogate risk bounds that quantify that convergence rate. We prove a linear surrogate risk bound for adversarially consistent losses (Theorem 9). We establish a distribution-dependent surrogate risk bound that applies whenever a loss is adversarially consistent for the data distribution (Theorem 11). |
| Researcher Affiliation | Academia | Natalie Frank EMAIL Department of Applied Mathematics University of Washington |
| Pseudocode | No | The paper focuses on theoretical derivations, proofs, and examples illustrating mathematical concepts. It does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing code, nor does it provide a link to a code repository or supplementary material containing code for the described methodology. The 'Open Review' link refers to a review forum, not a code base. |
| Open Datasets | Yes | Bhagoji et al. (2019) compute lower bounds on the adversarial classification risk for binary tasks, focusing on classifying digits 3 and 7 in MNIST under ℓ2 perturbations. Similar trends are observed on Fashion MNIST and CIFAR10. Dai et al. (2023) extend these bounds to the multiclass setting, though extending adversarial surrogate bounds beyond binary classification remains an open problem. When the optimal adversarial risk is non-zero, the adversarial Bayes classifier may not be unique up to degeneracy. Even without adversarial perturbations, datasets like MNIST and CIFAR10 contain inherently ambiguous examples. |
| Dataset Splits | No | The paper is theoretical and refers to prior work's experimental results on datasets like MNIST and CIFAR10. It does not describe any experiments conducted within this paper, and therefore, does not provide specific dataset split information for its own work. |
| Hardware Specification | No | The paper is theoretical, focusing on mathematical proofs and bounds, and does not report on experimental results. Therefore, no hardware specifications are mentioned. |
| Software Dependencies | No | The paper is theoretical and does not describe any specific software or programming environments used for empirical evaluation. Therefore, no software dependencies with version numbers are provided. |
| Experiment Setup | No | The paper is theoretical, focusing on mathematical proofs and derivations. It does not present any experimental results or detail an experimental setup with hyperparameters or training configurations. |