Research map / Adversarial robustness and security / Attacks on text and graphs
CoSPED: Consistent Soft Prompt Targeted Data Extraction and Defense
Zhuochen Yang, Kar Wai Fok, Vrizlynn L. L. Thing
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 Additive loss to MLE and Smooth loss (ablation) | CoSPED soft prompt · LM Extraction Benchmark D1 | ER50: +1.3% | Improved |
| add Dynamic loss to MLE loss (ablation) | CoSPED soft prompt · LM Extraction Benchmark D1 | ER50: +0.8% | Improved |
| use Focal loss with Smooth loss instead of MLE (ablation) | CoSPED soft prompt · LM Extraction Benchmark D1 | ER50: +0.6% | Improved |
| combine multiple extra losses rather than a single loss (ablation) | CoSPED soft prompt · LM Extraction Benchmark D1 | ER50: 61.2 -> 64.9 | Improved |
| Self Consistency decoding instead of beam search (ablation) | CoSPED soft prompt · LM Extraction Benchmark D1 | ER50: 64.5 -> 64.2 | No clear effect |
| Self Consistency decoding instead of greedy decoding (ablation) | CoSPED soft prompt · LM Extraction Benchmark D1 | ER50: 58.7 -> 64.2 | Improved |
| increase prefix length from 50 to 100 tokens (20-token soft prompt) (ablation) | CoSPED · LM Extraction Benchmark D2 | ER50: +23.8% | Improved |
| increase soft prompt length from 20 to 100 tokens (50-token prefix) (ablation) | CoSPED · LM Extraction Benchmark D2 | ER50: +8.8% | Improved |
| increase soft prompt length beyond 100 tokens (ablation) | CoSPED · LM Extraction Benchmark D2 | ER50: 64.5 -> 62.1 (50-token prefix, 100 -> 150 tokens) | No clear effect |
| use CoSPED over Ethicist across soft prompt and prefix lengths | GPT-Neo 1.3B · LM Extraction Benchmark D2 | ER50: 62.3 -> 65.2 (D1) | Improved |
| apply modified ROME editing as a defense | GPT-Neo 1.3B under CoSPED attack · 1000 test prefix-suffix pairs | ER50: 63.9 -> 0 | Improved |
| stop ROME editing at training PPL 10 instead of PPL 6 (ablation) | ROME defense on GPT-Neo 1.3B · LAMBADA | LB accuracy: 0.2598 -> 0.1247 | Hurt |
Closest papers by meaning
- Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM (NeurIPS 2025 · similarity 0.86)
- Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks (ICLR 2025 · similarity 0.86)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.86)
- Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy Leakage (AAAI 2025 · similarity 0.85)
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · similarity 0.85)
- Better Language Model Inversion by Compactly Representing Next-Token Distributions (NeurIPS 2025 · 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.
