Tab-MIA: A Benchmark for Membership Inference Attacks on Tabular Data

Tab-MIA is a benchmark dataset designed to evaluate the privacy risks of fine-tuning large language models (LLMs) on structured tabular data. It enables reproducible and systematic testing of Membership Inference Attacks (MIAs) across diverse datasets and six different serialization formats.

📋 Overview

• Datasets:

  • WTQ (WikiTableQuestions)
  • WikiSQL
  • TabFact
  • Adult Census
  • California Housing

• Encodings:

  • json
  • html
  • markdown
  • key-value-pair
  • key-is-value
  • line-sep

Each table is serialized into one of these formats and labeled with label=1 (member) or label=0 (non-member) for evaluating MIA methods.

📁 File Format

All files are in JSONL format. Each line has:

{
  "input": "serialized table string",
  "label": 1
}