Research map / Federated learning and privacy / Privacy attacks and secure learning
Anti-adversarial Learning: Desensitizing Prompts for Large Language Model
Xuan Li, Zhe Yin, Xiaodong Gu, Beijun Shen
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 |
|---|---|---|---|
| use gradient-based candidate selection instead of top-1 prediction (ablation) | PromptObfus (k=0.1) · AG News | MTI Top1 (lower is better): 35.06 -> 24.68 | Improved |
| use gradient-based candidate selection instead of random selection (ablation) | PromptObfus (k=0.1) · AG News | accuracy: 83.75 -> 85.25 | Improved |
| use gradient-based candidate selection instead of <MASK> insertion (ablation) | PromptObfus (k=0.1) · AG News | accuracy: 83.25 -> 85.25 | Improved |
| use TF-IDF rare-word masking instead of random masking (ablation) | PromptObfus (k=0.1) · AG News | accuracy: 84.25 -> 85.25 | Improved |
| use TF-IDF rare-word masking instead of random masking (ablation) | PromptObfus (k=0.1) · AG News | MTI Top1 (lower is better): 24.33 -> 24.68 | No clear effect |
| increase obfuscation ratio k from 0.1 to 0.3 (ablation) | PromptObfus · AG News | accuracy: 85.25 -> 83.75 | Hurt |
| increase obfuscation ratio k from 0.1 to 0.3 (ablation) | PromptObfus · AG News | EI Top1 (lower is better): 61.86 -> 49.65 | Improved |
| increase obfuscation ratio k from 0.1 to 0.3 (ablation) | PromptObfus · PersonalPortrait | PI occupation success rate (lower is better): 45.75 -> 17.25 | Improved |
| use larger LoRA-tuned surrogate (Llama-2-7B) instead of BART-large (ablation) | PromptObfus surrogate · SST sentiment | accuracy: 86.67 -> 84.80 | Hurt |
| change surrogate architecture or scale (ablation) | PromptObfus surrogate · SST sentiment | MTI Top1 / EI Top1: MTI 42.34-43.99; EI 81.71-81.80 | No clear effect |
| use PromptObfus instead of SANTEXT+ for utility | PromptObfus (k=0.1) · AG News | accuracy: 55.25 -> 85.25 | Improved |
| use PromptObfus instead of SANTEXT+ for EI attack resistance | PromptObfus (k=0.1) · AG News | EI Top1 (lower is better): 49.09 -> 61.86 | Hurt |
Closest papers by meaning
- DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer (ICLR 2024 · similarity 0.89)
- AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs (ICML 2025 · similarity 0.88)
- SentinelLMs: Encrypted Input Adaptation and Fine-Tuning of Language Models for Private and Secure Inference (AAAI 2024 · similarity 0.87)
- Privacy-Preserving Personalized Federated Prompt Learning for Multimodal Large Language Models (ICLR 2025 · similarity 0.87)
- Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs (ICML 2025 · similarity 0.87)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · 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.
