Can Prompt-Tuning Language Models Ensure Privacy?

The researchers propose a privacy attack framework targeting prompt-tuning that infers users' private information by exploiting prompt modules carrying user-specific signals. The paper first describes a real email service pipeline that provides customized outputs to different users through prompt-tuning, then conducts a comprehensive privacy evaluation on this target pipeline, with results showing that prompt-tuning has potential privacy leakage and that the proposed attack is effective.