Research map / Large language models / Language, safety and interpretability
Membership Inference Attack Against Large Language Model-Based Recommendation Systems: A New Distillation-Based Paradigm
Cuihong Li, Xiaowen Huang, Chuanhuan Yin, Jitao Sang
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 distilled reference model with fused features instead of shadow model | MIA on LLM-based recommenders (T5) · MovieLens | AUC %: 67.74 -> 90.14 | Improved |
| use distilled reference model with fused features instead of shadow model | MIA on LLM-based recommenders (T5) · Delicious | accuracy %: 51.35 -> 66.36 | Improved |
| replace logistic regression attack model with BERT (ablation) | MIA · Last.FM and MovieLens (small data) | accuracy %: 65.09 -> 59.09 (Last.FM); 85.00 -> 70.00 (MovieLens) | Hurt |
| replace logistic regression with BERT on large data (ablation) | MIA · Delicious | accuracy %: 65.27 -> 66.71, training time 16.17s -> 691.36s | No clear effect |
| use fused feature instead of individual features (ablation) | MIA · MovieLens | AUC %: 84.36 (loss only) -> 90.84 | Improved |
| use fused feature instead of loss-only on Delicious (ablation) | MIA · Delicious | f1 %: 73.12 (loss only) -> 72.28 | Hurt |
| keep loss in fused feature (ablation) | MIA · MovieLens | AUC %: 75.42 -> 90.84 | Improved |
| keep vector feature in fused feature (ablation) | MIA · Delicious | AUC %: 68.29 -> 67.55 | No clear effect |
| weight features in fusion (ablation) | MIA · MovieLens | accuracy %: 80.63 -> 86.00 | Improved |
| compress MLP-lifted features back to 512D (ablation) | MIA · MovieLens | accuracy %: 83.75 -> 76.25 | Hurt |
| use student with architecture matching the teacher (ablation) | MIA · MovieLens (T5 teacher) | AUC %: T5 student 90.42 vs GPT-2 student 63.38 | Improved |
| vary distillation weight alpha (ablation) | MIA · four datasets | attack performance | Mixed |
Closest papers by meaning
- Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · similarity 0.89)
- DF-MIA: A Distribution-Free Membership Inference Attack on Fine-Tuned Large Language Models (AAAI 2025 · similarity 0.88)
- Exploring the limits of strong membership inference attacks on large language models (NeurIPS 2025 · similarity 0.88)
- MIA-Tuner: Adapting Large Language Models as Pre-training Text Detector (AAAI 2025 · similarity 0.87)
- Active Large Language Model-Based Knowledge Distillation for Session-Based Recommendation (AAAI 2025 · similarity 0.87)
- Membership Inference Attacks against Large Vision-Language Models (NeurIPS 2024 · similarity 0.86)
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.
