Research map / Adversarial robustness and security / Certified and empirical robustness
Breaking the Adversarial Robustness-Performance Trade-off in Text Classification via Manifold Purification
Chenhao Dang, Jing Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add negative log-likelihood loss (density estimation) to MC2F training (ablation) | MC2F · AGNews TextFooler | accuracy under attack (Aua%): 32.6 -> 53.8 | Improved |
| add topological regularization loss to MC2F training (ablation) | MC2F · AGNews TextFooler | accuracy under attack (Aua%): 32.9 -> 53.8 | Improved |
| add causal and semantic (Fisher-Rao) loss to MC2F training (ablation) | MC2F · AGNews TextFooler | accuracy under attack (Aua%): 48.6 -> 53.8 | Improved |
| add topological regularization loss to MC2F training (ablation) | MC2F · AGNews clean accuracy | Clean%: 93.41 -> 95.13 | Improved |
| add causal loss to MC2F training (ablation) | MC2F · AGNews clean accuracy | Clean%: 94.76 -> 95.13 | No clear effect |
| replace fine-tuned BERT with MC2F purification defense | BERT-base · AGNews TextFooler | accuracy under attack (Aua%): 14.7 -> 53.8 | Improved |
| use MC2F instead of Subspace Defense (SD) | BERT-base · AGNews BERT-Attack | accuracy under attack (Aua%): 38.6 -> 45.3 | Improved |
| use MC2F instead of Subspace Defense (SD) | BERT-base · YELP clean test set | Clean%: 93.45 -> 95.26 | Improved |
| use MC2F instead of Subspace Defense (SD) | BERT-base · YELP BERT-Attack | average queries per successful attack: 320.7 -> 586.4 | Improved |
| use MC2F instead of standard fine-tuning | BERT-base · SST-2 clean test set | Clean%: 92.71 -> 92.71 | No clear effect |
Closest papers by meaning
- Adversarial Purification with the Manifold Hypothesis (AAAI 2024 · similarity 0.87)
- Adversarial Purification by Consistency-Aware Latent Space Optimization on Data Manifolds (TPAMI 2026 · similarity 0.86)
- Robust Graph Condensation via Classification Complexity Mitigation (NeurIPS 2025 · similarity 0.85)
- Enhancing the Adversarial Robustness via Manifold Projection (AAAI 2025 · similarity 0.85)
- STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing from Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.84)
- Explaining the role of Intrinsic Dimensionality in Adversarial Training (ICML 2025 · similarity 0.84)
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.
