Scholarlink

In today’s data-driven landscape, the surge in deep learning models raises concerns about privacy vulnerabilities. Attacks on deep learning models pose a significant threat, exploiting model outputs to determine if specific data points were in the training set. This ends up in risking privacy breaches. This risk is exacerbated as models become more prone to overfitting, inadvertently revealing sensitive information. To prevent these challenges, this paper proposes an innovative approach focused on generating synthetic data to augment training datasets. By expanding the data volume, susceptibility to overfitting and attacks can be greatly reduced. Central to this solution is Layer-wise Relevance Propagation (LRP), which identifies and quantifies feature relevance. Leveraging LRP insights, the method incorporates adaptive Differential Privacy Generative Adversarial Networks (DP-GANs) to produce synthetic data with enhanced privacy protections. By using DP-GANs with adaptive noise addition using LRP, this approach ensures that sensitive information remains secure while enriching the training dataset. The result is a robust framework for augmenting datasets, preserving privacy, and strengthening deep learning models with improved accuracy.

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

Elevating Privacy: A Differential Privacy Infused Approach to GAN for Robust Data Synthesis for Deep Learning Models

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • S. Sangeetha,
  • E. Shriaarthy,
  • N. Rithvik Pranao,
  • J. M. Suriya Priya,
  • L. Yogeswari,
  • S. Gokul,
  • S. S. Sri Nandha

摘要

In today’s data-driven landscape, the surge in deep learning models raises concerns about privacy vulnerabilities. Attacks on deep learning models pose a significant threat, exploiting model outputs to determine if specific data points were in the training set. This ends up in risking privacy breaches. This risk is exacerbated as models become more prone to overfitting, inadvertently revealing sensitive information. To prevent these challenges, this paper proposes an innovative approach focused on generating synthetic data to augment training datasets. By expanding the data volume, susceptibility to overfitting and attacks can be greatly reduced. Central to this solution is Layer-wise Relevance Propagation (LRP), which identifies and quantifies feature relevance. Leveraging LRP insights, the method incorporates adaptive Differential Privacy Generative Adversarial Networks (DP-GANs) to produce synthetic data with enhanced privacy protections. By using DP-GANs with adaptive noise addition using LRP, this approach ensures that sensitive information remains secure while enriching the training dataset. The result is a robust framework for augmenting datasets, preserving privacy, and strengthening deep learning models with improved accuracy.