Research map / Large language models / Language, safety and interpretability
Tab-MIA: A Benchmark Dataset for Membership Inference Attacks on Tabular Data in LLMs
Eyal German, Sagiv Antebi, Daniel Samira, Asaf Shabtai, Yuval Elovici
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 |
|---|---|---|---|
| fine-tune for 3 epochs instead of 1 (ablation) | Min-K++ 20.0% attack on LLaMA-3.1 8B · WTQ, Line-Separated | AUROC (%): 61.60 -> 93.60 | Hurt |
| fine-tune for 3 epochs instead of 1 (ablation) | Min-K++ 20.0% attack on Mistral 7B · WTQ, Line-Separated | AUROC (%): 69.70 -> 97.70 | Hurt |
| fine-tune for 3 epochs instead of 1 (ablation) | Min-K++ 20.0% attack on LLaMA-3.1 8B · Adult, Line-Separated | AUROC (%): 55.10 -> 71.10 | Hurt |
| use Min-K++ 20.0% instead of LOSS (PPL) | MIA on LLaMA-3.1 8B · WikiSQL, Line-Separated, 3 epochs | AUROC (%): 76.50 -> 88.90 | Improved |
| use Min-K++ 20.0% instead of Min-K 20.0% | MIA on LLaMA-3.1 8B · WikiSQL, Line-Separated, 3 epochs | AUROC (%): 78.10 -> 88.90 | Improved |
| use larger model (LLaMA-3.1 8B vs LLaMA-3.2 3B) | MIA target model · WTQ, Line-Separated, 3 epochs | AUROC (%): 89.70 -> 93.60 | Hurt |
| encode with Line-Separated instead of JSON (ablation) | Mistral 7B · WTQ | AUROC (%): 82.70 -> 97.70 | Hurt |
| encode with Key-Value Pair instead of JSON (ablation) | Mistral 7B · California Housing | AUROC (%): 54.50 -> 92.60 | Hurt |
| encode with Key-Value Pair instead of HTML (ablation) | Mistral 7B · California Housing | AUROC (%): 50.60 -> 92.60 | Hurt |
| attack with mismatched format (Key-Value) vs matched Markdown | Gemma-3 4B · TabFact cross-format | AUROC (%): 85.2 -> 68.9 | Hurt |
| attack with mismatched format (Line-Separated) vs matched Markdown | Gemma-3 4B · TabFact cross-format | AUROC (%): 85.2 -> 69.4 | Hurt |
| measure pretrained model with no fine-tuning | Llama-3.1 8B, Key-Value Pair · WTQ pretrained | AUROC (%): 72.0 | Mixed |
Closest papers by meaning
- MIA-Tuner: Adapting Large Language Models as Pre-training Text Detector (AAAI 2025 · similarity 0.90)
- Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · similarity 0.90)
- LLM Dataset Inference: Did you train on my dataset? (NeurIPS 2024 · similarity 0.89)
- Exploring the limits of strong membership inference attacks on large language models (NeurIPS 2025 · similarity 0.89)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.88)
- The Canary’s Echo: Auditing Privacy Risks of LLM-Generated Synthetic Text (ICML 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.
