Scholarlink

<p>The proliferation of digital financial platforms has led to the generation of large volumes of sensitive financial data, posing significant challenges for secure and reliable credit risk prediction. Conventional privacy-preserving machine learning approaches often incur high computational overhead, particularly when differential privacy is integrated into deep neural network training. To address these challenges, this paper proposes a differential privacy-enabled privacy-preserving neural network (DP-PPNN) framework designed to ensure data confidentiality throughout the prediction pipeline. The proposed framework incorporates differential privacy through Laplace and Gaussian noise mechanisms and systematically evaluates their impact on model performance. Privacy preservation is enforced at multiple stages, including data transformation, gradient updates during training, and output prediction. Specifically, noise injection is applied to input data, stochastic optimization steps via gradient clipping and perturbation, and final model outputs to ensure robust privacy guarantees. Experimental results on a real-world bank loan dataset demonstrate that the Laplace noise-based approach achieves better performance in terms of accuracy and loss compared to the Gaussian mechanism at a privacy budget of ε = 0.2, across batch sizes of 1, 32, and 64. The findings highlight the effectiveness of the proposed DP-PPNN framework in balancing privacy and predictive performance, providing a practical solution for secure credit risk assessment in financial systems.</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

Differential privacy-enabled neural networks for secure credit risk forecasting in banking

  • 期刊论文

  • 发表日期：

    2026年6月21日

查看全文
  • Vankamamidi S. Naresh,
  • Sivaranjani Reddi,
  • M. Thamarai

摘要

The proliferation of digital financial platforms has led to the generation of large volumes of sensitive financial data, posing significant challenges for secure and reliable credit risk prediction. Conventional privacy-preserving machine learning approaches often incur high computational overhead, particularly when differential privacy is integrated into deep neural network training. To address these challenges, this paper proposes a differential privacy-enabled privacy-preserving neural network (DP-PPNN) framework designed to ensure data confidentiality throughout the prediction pipeline. The proposed framework incorporates differential privacy through Laplace and Gaussian noise mechanisms and systematically evaluates their impact on model performance. Privacy preservation is enforced at multiple stages, including data transformation, gradient updates during training, and output prediction. Specifically, noise injection is applied to input data, stochastic optimization steps via gradient clipping and perturbation, and final model outputs to ensure robust privacy guarantees. Experimental results on a real-world bank loan dataset demonstrate that the Laplace noise-based approach achieves better performance in terms of accuracy and loss compared to the Gaussian mechanism at a privacy budget of ε = 0.2, across batch sizes of 1, 32, and 64. The findings highlight the effectiveness of the proposed DP-PPNN framework in balancing privacy and predictive performance, providing a practical solution for secure credit risk assessment in financial systems.