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

On Adversarial Training without Perturbing all Examples

ICLR 2024 · Adversarial robustness and security › Certified and empirical robustness · approach: Regularized defenses against attacks

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.

Explores adversarial training where adversarial examples are constructed for only a subset of training examples, with the rest trained on clean data, to reduce cost.

Problem

Adversarial training is the standard defense against adversarial examples but requires a multi-step attack on every training example, which is expensive.

Approach

Split the training set into subsets A and B, train on both, but construct adversarial examples only on subset A while B is used as clean examples, reducing the number of examples that need multi-step attacks.

Experimental setup

Builds on: adversarial training

Closest papers by meaning

• Improving Adversarial Robustness Requires Revisiting Misclassified Examples (ICLR 2020 · similarity 0.93)
• Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.92)
• Adversarial Machine Learning at Scale (ICLR 2017 · similarity 0.92)
• Are Adversarial Examples Created Equal? A Learnable Weighted Minimax Risk for Robustness under Non-uniform Attacks (AAAI 2021 · similarity 0.91)
• Squeeze Training for Adversarial Robustness (ICLR 2023 · similarity 0.91)
• A Closer Look at Curriculum Adversarial Training: From an Online Perspective (AAAI 2024 · similarity 0.90)