Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Training for Improving Model Robustness? Look at Both Prediction and Interpretation
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 |
|---|---|---|---|
| use FLAT rather than traditional adversarial training | LSTM · SST2 test (Textfooler) | after-attack accuracy: 12.36 -> 17.76 | Helped |
| use FLAT rather than traditional adversarial training | CNN · SST2 test (Textfooler) | after-attack accuracy: 2.53 -> 37.07 | Helped |
| use FLAT rather than traditional adversarial training | BERT · SST2 test (Textfooler) | after-attack accuracy: 5.60 -> 12.41 | Helped |
| use FLAT rather than traditional adversarial training | DeBERTa · SST2 test (Textfooler) | after-attack accuracy: 7.96 -> 11.59 | Helped |
| add importance-score regularization for replaced words and their substitutions (ablation) | FLAT with β=0.1, γ=0 on LSTM · SST2 test (Textfooler) | after-attack accuracy: 9.99 -> 17.76 | Helped |
| add word-mask entropy regularization (ablation) | FLAT with β=0, γ=0.001 on LSTM · SST2 test (Textfooler) | after-attack accuracy: 8.18 -> 17.76 | Helped |
| add both word-mask entropy and importance-score regularizations (ablation) | FLAT with β=0, γ=0 on LSTM · SST2 test (Textfooler) | after-attack accuracy: 8.35 -> 17.76 | Helped |
| add importance-score regularization for replaced words and their substitutions (ablation) | FLAT with β=0.1, γ=0 on LSTM · SST2 test (Textfooler) | standard accuracy: 83.96 -> 84.79 | Helped |
| add word-mask entropy regularization (ablation) | FLAT with β=0, γ=0.001 on LSTM · SST2 test (Textfooler) | standard accuracy: 84.34 -> 84.79 | Helped |
| replace learned word masks with predefined similarity-based word groups (ablation) | CNN on AG · AG test (Textfooler) | after-attack accuracy: 10: 1.20; 20: 1.92; 30: 1.40; 40: 1.46; 50: 1.38 | Hurt |
Closest papers by meaning
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Bag of Tricks for Adversarial Training (ICLR 2021 · similarity 0.89)
- Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models (NeurIPS 2021 · similarity 0.89)
- Scaling Trends in Language Model Robustness (ICML 2025 · similarity 0.89)
- Towards Robust Knowledge Unlearning: An Adversarial Framework for Assessing and Improving Unlearning Robustness in Large Language Models (AAAI 2025 · similarity 0.88)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 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.