Research map / Adversarial robustness and security / Certified and empirical robustness
Understanding and Improving Fast Adversarial Training against $l_0$ Bounded Perturbations
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.
Studies fast (single-step) adversarial training against sparse l0-bounded perturbations, diagnosing catastrophic overfitting in this setting and proposing an improved fast AT method.
Problem
Fast adversarial training against l0-bounded sparse perturbations suffers from catastrophic overfitting, which existing fast AT methods designed for l_p norms (p>=1) do not properly address.
Approach
Identifies the specific mechanism of catastrophic overfitting for 1-step l0 attacks and proposes a fast AT modification tailored to l0 perturbations, rather than transferring l_p-norm fixes directly.
Surprising findings
- Catastrophic overfitting in l0 adversarial training cannot be properly addressed by existing fast AT methods developed for other l_p norms (p>=1).
Experimental setup
Baselines: existing fast adversarial training methods for l_p norms
Builds on: fast adversarial training catastrophic overfitting analysis
Closest papers by meaning
- Towards Stable and Efficient Adversarial Training against $l_1$ Bounded Adversarial Attacks (ICML 2023 · similarity 0.92)
- Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations (ICML 2024 · similarity 0.91)
- Fast Minimum-norm Adversarial Attacks through Adaptive Norm Constraints (NeurIPS 2021 · similarity 0.90)
- $\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples (ICLR 2025 · similarity 0.89)
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 2024 · similarity 0.88)
- Layer-Aware Analysis of Catastrophic Overfitting: Revealing the Pseudo-Robust Shortcut Dependency (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.