Are Large Pre-Trained Language Models Leaking Your Personal Information?
This work addresses privacy risks for users of PLMs, though it is incremental in analyzing existing vulnerabilities.
The paper investigates whether large pre-trained language models (PLMs) leak personal information, such as email addresses, by querying them with contexts or prompts containing owner names, finding that PLMs do leak due to memorization but with low risk from attackers due to weak association.
Are Large Pre-Trained Language Models Leaking Your Personal Information? In this paper, we analyze whether Pre-Trained Language Models (PLMs) are prone to leaking personal information. Specifically, we query PLMs for email addresses with contexts of the email address or prompts containing the owner's name. We find that PLMs do leak personal information due to memorization. However, since the models are weak at association, the risk of specific personal information being extracted by attackers is low. We hope this work could help the community to better understand the privacy risk of PLMs and bring new insights to make PLMs safe.