Scholarlink

<p>With the increasing prevalence of machine learning applications in financial data analysis, safeguarding the privacy of customers’ personal financial information during credit risk assessment is paramount. Despite ongoing research efforts in this field, establishing robust privacy-preserving credit risk prediction systems capable of mitigating diverse privacy attacks remains a formidable challenge. To address this issue, we propose a Privacy-Preserving Credit Risk Analysis (PPCRA) framework that leverages homomorphic encryption-aware Machine Learning (ML) on encrypted data. Various ML and Privacy-Preserving Machine Learning (PPML) models were built using the TenSEAL and Concrete ML on datasets from Germany, Taiwan, Japan, and Australia. When comparing PPML with traditional ML models, it is evident that PPML achieves a considerable level of privacy preservation with only minimal loss in accuracy. The experimental results indicate that Privacy-Preserving Logistic Regression (PPLR) outperformed other PPML models. Furthermore, the security analysis demonstrates that the proposed system effectively withstands multiple privacy threats, including poisoning, membership inference, evasion, model extraction, and model inversion attacks, across various stages of the machine learning lifecycle.</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

Privacy-preserving machine learning techniques based on homomorphic encryption for credit risk analysis

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • V. V. L. Divakar Allavarpu,
  • Vankamamidi S. Naresh,
  • A. Krishna Mohan

摘要

With the increasing prevalence of machine learning applications in financial data analysis, safeguarding the privacy of customers’ personal financial information during credit risk assessment is paramount. Despite ongoing research efforts in this field, establishing robust privacy-preserving credit risk prediction systems capable of mitigating diverse privacy attacks remains a formidable challenge. To address this issue, we propose a Privacy-Preserving Credit Risk Analysis (PPCRA) framework that leverages homomorphic encryption-aware Machine Learning (ML) on encrypted data. Various ML and Privacy-Preserving Machine Learning (PPML) models were built using the TenSEAL and Concrete ML on datasets from Germany, Taiwan, Japan, and Australia. When comparing PPML with traditional ML models, it is evident that PPML achieves a considerable level of privacy preservation with only minimal loss in accuracy. The experimental results indicate that Privacy-Preserving Logistic Regression (PPLR) outperformed other PPML models. Furthermore, the security analysis demonstrates that the proposed system effectively withstands multiple privacy threats, including poisoning, membership inference, evasion, model extraction, and model inversion attacks, across various stages of the machine learning lifecycle.