Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Can We Infer Confidential Properties of Training Data from LLMs?

Authors: Pengrun Huang, Chhavi Yadav, Kamalika Chaudhuri, Ruihan Wu

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations across multiple pretrained LLMs show the success of our attacks, revealing a previously unrecognized vulnerability in LLMs. We release our code at github.com/Pengrun H/Property_inference_attack_LLM.
Researcher Affiliation Academia Pengrun Huang, Chhavi Yadav, Kamalika Chaudhuri , Ruihan Wu University of California, San Diego EMAIL
Pseudocode No The paper describes the methods and attacks in prose, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps.
Open Source Code Yes We release our code at github.com/Pengrun H/Property_inference_attack_LLM.
Open Datasets Yes Built on the Chat Doctor dataset, our benchmark includes a range of property types and task configurations. 1https://huggingface.co/datasets/Pengrun/Prop Infer_dataset
Dataset Splits Yes We use 15,000 samples to train the target models and the remaining 14,791 as auxiliary data for evaluating attacks in the grey-box setting. For medical diagnosis attributes, we use the original training split of the Chat Doctor dataset with size 50, 000 for training the target models
Hardware Specification Yes All experiments are conducted on NVIDIA RTX 6000 Ada GPU.
Software Dependencies No The paper mentions software like XGBoost [7] and scikit-learn library [23], but does not provide specific version numbers for these or other key software components.
Experiment Setup Yes We used the Lo RA [15] method for fine-tuning with a learning rate of 1e 4, dropout rate of 0.05, Lo RA rank of 128 and 5 epochs.