Scholarlink

<p>The given paper introduces an optimized differential private stochastic gradient descent (DP-SGD) approach to simulate membership inference attacks in the given neural network. Membership inference attacks are a powerful threat to individual privacy in datasets, which allows adversaries to use trained models and identify whether a specific data sample is in the model's training set. The proposed approach is implemented and validated based on performance metrics such as gradient clipping and privacy budget by taking two datasets- CIFAR 10 and MNIST into consideration. The results show that DP-SGD can be employed to decrease the membership attack success rate quite drastically thereby enhancing the trained neural network model accuracy. In addition to this, the proposed approach is validated and compared with existing studies based on accuracy (%). DP-SGD approach attains the highest accuracy (93.85%) thereby outperforming the existing recent studies.</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

An optimized differential private stochastic gradient descent (DP-SGD) approach to combat membership inference attacks in neural networks

  • 期刊论文

  • 发表日期：

    2025年2月23日

查看全文
  • Emad Elabd,
  • Waleed M. Ead,
  • Shalu

摘要

The given paper introduces an optimized differential private stochastic gradient descent (DP-SGD) approach to simulate membership inference attacks in the given neural network. Membership inference attacks are a powerful threat to individual privacy in datasets, which allows adversaries to use trained models and identify whether a specific data sample is in the model's training set. The proposed approach is implemented and validated based on performance metrics such as gradient clipping and privacy budget by taking two datasets- CIFAR 10 and MNIST into consideration. The results show that DP-SGD can be employed to decrease the membership attack success rate quite drastically thereby enhancing the trained neural network model accuracy. In addition to this, the proposed approach is validated and compared with existing studies based on accuracy (%). DP-SGD approach attains the highest accuracy (93.85%) thereby outperforming the existing recent studies.