Scholarlink

The representation encoding of training data by models frequently captures sensitive individual attributes, such as gender and age, thereby giving rise to privacy concerns. For most existing methods, the utilization of target classifier and attack classifier outputs for loss computation disrupts the direction of model updates, increasing the risk of privacy leakage in the encoded representation vectors produced by the model. To address this issue, a Differential Privacy Adversarial Learning with Diffusion Model (DPALDM) is proposed. The minimum leakage rates on the three datasets are reduced to 54.82%, 52.8%, and 16.39%, respectively, indicating that the proposed method effectively safeguards sensitive attributes in encoded representation vectors and enhances privacy protection effectiveness.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Differential Privacy Adversarial Learning with Diffusion Model for the Generation of Privacy-Protected Text Representations

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yeli Guan

摘要

The representation encoding of training data by models frequently captures sensitive individual attributes, such as gender and age, thereby giving rise to privacy concerns. For most existing methods, the utilization of target classifier and attack classifier outputs for loss computation disrupts the direction of model updates, increasing the risk of privacy leakage in the encoded representation vectors produced by the model. To address this issue, a Differential Privacy Adversarial Learning with Diffusion Model (DPALDM) is proposed. The minimum leakage rates on the three datasets are reduced to 54.82%, 52.8%, and 16.39%, respectively, indicating that the proposed method effectively safeguards sensitive attributes in encoded representation vectors and enhances privacy protection effectiveness.