Research map / Adversarial robustness and security / Certified and empirical robustness

Algorithmic Stability Based Generalization Bounds for Adversarial Training

ICLR 2025 · Adversarial robustness and security › Certified and empirical robustness · approach: Provable robust learning guarantees

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proves algorithmic-stability generalization upper bounds for adversarial training, governed by expansiveness parameters of the adversarial perturbations used in training and evaluation.

Problem

Generalization behavior of adversarial training lacks tight theoretical bounds; existing analyses rely on artificial assumptions.

Approach

Uses an algorithmic stability analysis in which the expansiveness of adversarial perturbations (used during training and evaluation) controls both the vanishing rate and the scaling constant of the generalization error, without artificial assumptions.

Closest papers by meaning

• On the Generalization Analysis of Adversarial Learning (ICML 2022 · similarity 0.90)
• Stability Analysis and Generalization Bounds of Adversarial Training (NeurIPS 2022 · similarity 0.90)
• Attacks Which Do Not Kill Training Make Adversarial Learning Stronger (ICML 2020 · similarity 0.90)
• Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation (NeurIPS 2024 · similarity 0.90)
• Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations (ICML 2024 · similarity 0.88)
• Uniformly Stable Algorithms for Adversarial Training and Beyond (ICML 2024 · similarity 0.88)