Scholarlink

<p>With the rapid development of social media environments, the usage of digital text documents has significantly increased, leading to the leakage of crucial sensitive information. Hence, several natural language processing techniques are developed for identifying sensitive data in textual documents but prone to class imbalance processing. Also, some methods struggle to capture contextual hidden information, temporal correlation as well as the semantic information that existed in the textual documents which are crucial for discriminating sensitive and non-sensitive entities. A novel framework named PrivacyTextNet to accurately detect sensitive information in textual documents is proposed. At first, the dual branch sequential variational autoencoder (DBSVAE) model is developed to handle the class imbalance problem that generates instances for sensitive entities as the occurrence of the sensitive one is very much smaller in the datasets. In the second phase, the sensitive information detection is performed by the PrivacyTextNet model. The PrivacyTextNet model comprises six layers to learn the complex characteristics of the textual documents. The bidirectional encoder representation transformer is initially applied for generating word embedding representations, and the contextual hidden information is learned by applying newly invented bidirectional gated recurrent-based conditional random field layer. The generated output vector is finally fed into the classification layer for categorizing information into sensitive and non-sensitive. To validate the proposed method’s performance, comprehensive experiments are conducted and error analysis is performed. The proposed method attained a detection accuracy of 98.57%, a false positive rate (FPR) reduced from 3.49–7.22% (baseline) to 1.43%, a false negative rate (FGR) reduced from 5.82–7.73% (baseline) to 3.17%, and a computational time of 2.3&#xa0;s, respectively. The PrivacyTextNet model performs reasonably well in recognizing sensitive information with the relatively lower false predictions. Overall, the proposed method provided outstanding results in sensitive information detection.</p>

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

Enhancing online privacy: a novel dual autoencoder framework for detecting sensitive information

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Janani Muralitharan,
  • Arumugam Chandrasekar

摘要

With the rapid development of social media environments, the usage of digital text documents has significantly increased, leading to the leakage of crucial sensitive information. Hence, several natural language processing techniques are developed for identifying sensitive data in textual documents but prone to class imbalance processing. Also, some methods struggle to capture contextual hidden information, temporal correlation as well as the semantic information that existed in the textual documents which are crucial for discriminating sensitive and non-sensitive entities. A novel framework named PrivacyTextNet to accurately detect sensitive information in textual documents is proposed. At first, the dual branch sequential variational autoencoder (DBSVAE) model is developed to handle the class imbalance problem that generates instances for sensitive entities as the occurrence of the sensitive one is very much smaller in the datasets. In the second phase, the sensitive information detection is performed by the PrivacyTextNet model. The PrivacyTextNet model comprises six layers to learn the complex characteristics of the textual documents. The bidirectional encoder representation transformer is initially applied for generating word embedding representations, and the contextual hidden information is learned by applying newly invented bidirectional gated recurrent-based conditional random field layer. The generated output vector is finally fed into the classification layer for categorizing information into sensitive and non-sensitive. To validate the proposed method’s performance, comprehensive experiments are conducted and error analysis is performed. The proposed method attained a detection accuracy of 98.57%, a false positive rate (FPR) reduced from 3.49–7.22% (baseline) to 1.43%, a false negative rate (FGR) reduced from 5.82–7.73% (baseline) to 3.17%, and a computational time of 2.3 s, respectively. The PrivacyTextNet model performs reasonably well in recognizing sensitive information with the relatively lower false predictions. Overall, the proposed method provided outstanding results in sensitive information detection.