Reproduce “A Synthetic Dataset for Personal Attribute Inference” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hanna Yukhymenko, Robin Staab, Mark Vero, Martin Vechev
Researchers at ETH Zurich developed SynthPAI, the first open and privacy-preserving synthetic dataset for personal attribute inference (PAI) research, generated using a personalized LLM agent-based simulation framework. A human study found these synthetic comments are nearly indistinguishable from real ones, and experiments reproducing PAI findings on SynthPAI yield consistent results with real-world data.
Key insights
- Synthetic comments generated by the LLM agents are highly realistic, with human participants achieving only 51.9% accuracy when asked to distinguish them from real human-written texts.
- The PAI performance and trends observed across state-of-the-art LLMs on SynthPAI (e.g., GPT-4 achieving 76% accuracy) consistently align with those found using real-world datasets.
- Existing text anonymization methods remain largely ineffective against sophisticated LLM inference, as models retained approximately 50% accuracy on anonymized texts for certain attributes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.