Scholarlink

The primary challenge faced by contemporary enterprises as they integrate Large Language Models (LLMs) into their operational frameworks: ensuring the confidentiality of sensitive data, particularly Personal Identifiable Information (PII). Despite the advanced and well-secured database architectures employed in modern LLM tools, there is still a major threat for safeguard of PII in LLM databases. Our study aims to provide a solution to reduce the risks connected with PII data breaches by comparing all present data protection strategies and providing a combination of them. Using threshold of susceptibility of PII in existing anonymizer libraries we have given an analysis of thresholds impact on PII vulnerability to risks. This method will help to enhance the model development for PII data handling. Also, by this comparative analysis we can provide a solution of LLM data breach by suggesting an intermediate API serving as a protective layer before data enters in LLM databases, ensuring the security of PII data from data breach. Our study focuses on all three available methods to enhance data security: hashing, encryption and anonymization. With this analysis we aim to measure effectiveness of provided methods mentioned. Evaluating advantages and disadvantages of each technique, study will give vital insights essential for implementation of secure LLM applications.

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

Prevention of Cyberattacks on Personal Identifiable Information in Large Language Models Using Data Anonymization Techniques Along with Encryption

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Prashant Giridhar Shambharkar,
  • Lakshay Gupta,
  • Manav Yadav,
  • Lakshay Gupta

摘要

The primary challenge faced by contemporary enterprises as they integrate Large Language Models (LLMs) into their operational frameworks: ensuring the confidentiality of sensitive data, particularly Personal Identifiable Information (PII). Despite the advanced and well-secured database architectures employed in modern LLM tools, there is still a major threat for safeguard of PII in LLM databases. Our study aims to provide a solution to reduce the risks connected with PII data breaches by comparing all present data protection strategies and providing a combination of them. Using threshold of susceptibility of PII in existing anonymizer libraries we have given an analysis of thresholds impact on PII vulnerability to risks. This method will help to enhance the model development for PII data handling. Also, by this comparative analysis we can provide a solution of LLM data breach by suggesting an intermediate API serving as a protective layer before data enters in LLM databases, ensuring the security of PII data from data breach. Our study focuses on all three available methods to enhance data security: hashing, encryption and anonymization. With this analysis we aim to measure effectiveness of provided methods mentioned. Evaluating advantages and disadvantages of each technique, study will give vital insights essential for implementation of secure LLM applications.