Research map / Large language models / Language, safety and interpretability
Stop Tracking Me! Proactive Defense Against Attribute Inference Attack in LLMs
Dong Yan, Jian Liang, Ran He, Tieniu Tan
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 vocabulary extraction to TRACE (ablation) | TRACE · Synthetic, GPT-3.5 Top@1 | attribute inference accuracy (%): 31.43 -> 26.86 | Improved |
| add inference chain to TRACE (ablation) | TRACE · Synthetic, GPT-3.5 Top@1 | attribute inference accuracy (%): 31.62 -> 26.86 | Improved |
| add both VE and IC (full TRACE) versus neither (ablation) | TRACE · Synthetic, GPT-3.5 Top@1 | attribute inference accuracy (%): 37.14 -> 26.86 | Improved |
| add vocabulary extraction to TRACE (ablation) | TRACE · Synthetic, GPT-3.5 Top@3 | attribute inference accuracy (%): 58.86 -> 56.38 | Improved |
| set Top-K to 10 instead of 5 (ablation) | TRACE vocabulary extraction · Synthetic | attribute inference accuracy (%): 33.33 -> 31.62 | Improved |
| set Top-K to 10 instead of 15 (ablation) | TRACE vocabulary extraction · Synthetic | attribute inference accuracy (%): 34.10 -> 31.62 | Improved |
| add RPS suffix on top of TRACE anonymization | TRACE (GPT-3.5) · Synthetic, Llama2-7B-Chat | attribute inference accuracy (%): 22.48 -> 0.19 | Improved |
| replace FgAA with TRACE anonymization | FgAA (GPT-3.5) · Synthetic, GPT-4o target | attribute inference accuracy (%): 39.05 -> 28.38 | Improved |
| replace FgAA with TRACE using GPT-4o anonymizer | FgAA (GPT-4o) · Synthetic, GPT-4o target | attribute inference accuracy (%): 25.71 -> 26.67 | No clear effect |
| add MPS to TRACE for instruction-following model | TRACE · Qwen2.5-7B-Instruct | attribute inference accuracy (%): 22.67 -> 5.14 | Improved |
| MPS alone instead of FgAA | FgAA · Qwen2.5-7B-Instruct | attribute inference accuracy (%): 28.76 -> 10.29 | Improved |
| adaptive attacker drops last characters of RPS-defended text | RPS · Synthetic, Llama2-7B-Chat | attribute inference accuracy (%): 1.71 -> 27.43 | Hurt |
Closest papers by meaning
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.91)
- Self-Refining Language Model Anonymizers via Adversarial Distillation (NeurIPS 2025 · similarity 0.89)
- A Synthetic Dataset for Personal Attribute Inference (NeurIPS 2024 · similarity 0.89)
- Language Models are Advanced Anonymizers (ICLR 2025 · similarity 0.89)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.88)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.88)
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.
