Research map / Adversarial robustness and security / Attacks on text and graphs
Towards Robustness Against Natural Language Word Substitutions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace prior perturbation sets with convex-hull ASCC adversarial training | LSTM sentiment classifier · IMDB, Genetic attack | accuracy (%): 72.5 -> 79.0 | Helped |
| Replace prior perturbation sets with convex-hull ASCC adversarial training | LSTM sentiment classifier · IMDB, PWWS attack | accuracy (%): 66.7 -> 77.1 | Helped |
| Use ASCC adversarial training rather than axis-aligned bounds | BOW classifier · SNLI, Genetic attack | accuracy (%): 75.0 -> 76.3 | Helped |
| Use ASCC adversarial training rather than axis-aligned bounds | BOW classifier · SNLI, PWWS attack | accuracy (%): 72.1 -> 75.1 | Helped |
| Add entropy sparsity regularization (α=10) (ablation) | CNN ASCC-defense · IMDB, PWWS attack | accuracy (%): 61.6 -> 75.7 | Helped |
| Add entropy sparsity regularization (α=10) (ablation) | BOW ASCC-defense · SNLI, Genetic attack | accuracy (%): 73.4 -> 76.3 | Helped |
| Increase entropy regularization weight from α=10 to α=15 (ablation) | CNN ASCC-defense · IMDB, Genetic attack | accuracy (%): 78.2 -> 76.1 | Hurt |
| Increase entropy regularization weight from α=10 to α=15 (ablation) | BOW ASCC-defense · SNLI, PWWS attack | accuracy (%): 75.1 -> 73.3 | Hurt |
| Initialize with ASCC-defense-trained robust word vectors rather than GloVe | Normally trained LSTM · IMDB, Genetic attack | accuracy (%): 7.9 -> 73.4 | Helped |
| Apply ASCC-defense instead of standard fine-tuning | BERT · IMDB, Genetic attack | accuracy (%): 16.4 -> 70.2 | Helped |
| Use ASCC-defense instead of random augmentation | Text classifiers · IMDB and SNLI, Genetic attack | robust accuracy (%): +46.0% on IMDB; +7.8% on SNLI | Helped |
| Increase entropy regularization weight from α=10 to α=15 (ablation) | CNN ASCC-defense · IMDB, PWWS attack | accuracy (%): 75.7 -> 78.3 | Helped |
Closest papers by meaning
- REINFORCE Adversarial Attacks on Large Language Models: An Adaptive, Distributional, and Semantic Objective (ICML 2025 · similarity 0.87)
- Towards Robust Knowledge Unlearning: An Adversarial Framework for Assessing and Improving Unlearning Robustness in Large Language Models (AAAI 2025 · similarity 0.87)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.87)
- RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content (ICML 2024 · similarity 0.86)
- Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks (ICLR 2025 · similarity 0.86)
- On the Robustness of Verbal Confidence of LLMs in Adversarial Attacks (NeurIPS 2025 · similarity 0.86)
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.