Research map / Adversarial robustness and security / Attacks on text and graphs
Joint Character-Level Word Embedding and Adversarial Stability Training to Defend Adversarial Text
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 word embedding with char embedding (CharCNN-BLSTM vs Word-BLSTM) (ablation) | BLSTM classifier · AG gradient-based adversarial | accuracy (%): 53.41 -> 70.99 | Helped |
| replace word embedding with char embedding (ablation) | BLSTM classifier · AG natural adversarial | accuracy (%): 82.50 -> 87.88 | Helped |
| add similarity loss Ldiff (ablation) | CharCNN-BLSTM · AG gradient-based adversarial | accuracy (%): 70.99 -> 76.04 | Helped |
| add noisy-input loss Lnoisy (ablation) | CharCNN-BLSTM · AG gradient-based adversarial | accuracy (%): 70.99 -> 84.44 | Helped |
| combine Ldiff and Lnoisy (full AST) versus Lnoisy alone (ablation) | CharCNN-BLSTM · AG gradient-based adversarial | accuracy (%): 84.44 -> 84.34 | No clear effect |
| combine Ldiff and Lnoisy versus Lnoisy alone (ablation) | CharCNN-BLSTM · AG natural adversarial | accuracy (%): 89.39 -> 89.45 | Helped |
| set alpha (Lnoisy weight) to small values (ablation) | CharCNN-BLSTM-AST · AG | accuracy | Hurt |
| use CharBLSTM instead of CharCNN char embedding (ablation) | BLSTM-AST · mean drop over 5 datasets, gradient-based | accuracy drop (%): 8.51 -> 6.81 | Helped |
| char embedding with AST instead of spell correctors (ScRNN-10) | Word-BLSTM · mean drop over 5 datasets, natural | accuracy drop (%): 13.75 -> 5.17 | Helped |
Closest papers by meaning
- Adversarial Training for Improving Model Robustness? Look at Both Prediction and Interpretation (AAAI 2022 · similarity 0.88)
- UniT: A Unified Look at Certified Robust Training against Text Adversarial Perturbation (NeurIPS 2023 · similarity 0.88)
- TextGrad: Advancing Robustness Evaluation in NLP by Gradient-Driven Optimization (ICLR 2023 · similarity 0.87)
- Adversarial Training Methods for Semi-Supervised Text Classification (ICLR 2017 · similarity 0.87)
- Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models (NeurIPS 2021 · similarity 0.87)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.87)
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.