Research map / Adversarial robustness and security / Attacks on text and graphs
On Robust Prefix-Tuning for Text Classification
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 |
|---|---|---|---|
| Add test-time batch-level prefix tuning against canonical activation manifolds | Standard prefix-tuning · SST-2, AG's News, and SNLI adversarial attacks | accuracy: 16.64 -> 50.36 (PWWS), 1.92 -> 44.65 (VIPER), 31.58 -> 58.54 (SCPN), 8.84 -> 46.68 (BUG), 5.05 -> 85.72 (UAT) on SST-2; | Helped |
| Add test-time batch-level prefix tuning against canonical activation manifolds | Standard prefix-tuning · SST-2, AG's News, and SNLI clean data | accuracy: 92.48 -> 92.59 on SST-2; 86.42 -> 86.50 on AG's News; 72.48 -> 72.74 on SNLI | No clear effect |
| Use adversarial prefix-tuning instead of standard prefix-tuning | Prefix-tuning · SST-2, AG's News, and SNLI | accuracy: SST-2 clean: 92.48 -> 93.57; SST-2 UAT: 5.05 -> 4.88 | Mixed |
| Use a fixed test batch size of 1 rather than adaptive batching (ablation) | Robust prefix-tuning · SST-2 UAT | accuracy: 85.72 -> 73.09 | Hurt |
| Use test batch size 2 or 4 instead of batch size 1 (ablation) | Robust prefix-tuning · SST-2 UAT | accuracy: 73.09 -> 84.73 (batch size 2); 73.09 -> 86.66 (batch size 4) | Helped |
| Normalize batch activations before projection (ablation) | Robust prefix-tuning with batch size 2 or 4 · SST-2 UAT | accuracy: 78.20 -> 84.73 (batch size 2); 80.23 -> 86.66 (batch size 4) | Helped |
| Tune bottom layers rather than top layers (ablation) | Robust prefix-tuning · SST-2 development set attacks | accuracy | Helped |
| Tune bottom 3 layers rather than larger numbers of layers (ablation) | Robust prefix-tuning · SST-2 UAT | accuracy | Helped |
| Add PWWS adversarial data augmentation during training (ablation) | Standard prefix-tuning · SST-2, PWWS | accuracy: 16.64 -> 48.11 (20 epochs); 16.64 -> 55.08 (50 epochs); 16.64 -> 56.62 (100 epochs) | Helped |
| Add attack-specific adversarial data augmentation during training (ablation) | Standard prefix-tuning · SST-2 attacks other than augmentation attack | accuracy | Mixed |
| Add robust prefix-tuning to adversarial data augmentation models | SST-2 adversarially augmented prefix-tuning · SST-2 UAT | accuracy: 10.43 -> 70.07 (PWWS augmentation, 20 epochs); 3.13 -> 83.47 (PWWS augmentation, 50 epochs); 7.91 -> 86.38 (PWWS augment | Helped |
| Stop adversarial prefix-tuning at 20 epochs instead of 100 (ablation) | Adversarial prefix-tuning · SST-2 UAT | accuracy: 4.88 -> 6.97 (word-level); 8.95 -> 8.57 (sentence-level) | Mixed |
Closest papers by meaning
- RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content (ICML 2024 · similarity 0.88)
- Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks (ICLR 2025 · similarity 0.87)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.86)
- Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbation (NeurIPS 2025 · similarity 0.86)
- Representation Noising: A Defence Mechanism Against Harmful Finetuning (NeurIPS 2024 · similarity 0.86)
- Fight Back Against Jailbreaking via Prompt Adversarial Tuning (NeurIPS 2024 · similarity 0.85)
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.