<ul data-eligibleForWebStory="true"><li>Large language models (LLMs) are being fine-tuned on domain-specific datasets, which may contain sensitive and confidential information like patient demographics.</li><li>A new benchmark task called PropInfer is introduced to assess property inference in LLMs under question-answering and chat-completion fine-tuning paradigms.</li><li>PropInfer is built on the ChatDoctor dataset and includes various property types and task configurations.</li><li>The study explores prompt-based generation and shadow-model attacks to evaluate property inference in LLMs.</li><li>Empirical evaluations on multiple pretrained LLMs demonstrate the success of these attacks, highlighting a vulnerability in LLMs when it comes to inferring confidential properties of training data.</li></ul>

Can We Infer Confidential Properties of Training Data from LLMs?

Discover more