| 基于深度学习的动态差分隐私保护算法 |
 点此下载全文 |
| 引用本文：何佳阳.基于深度学习的动态差分隐私保护算法[J].计算技术与自动化,2024,(4):117-122 |
| 摘要点击次数: 451 |
| 全文下载次数: 0 |
|
|
| 中文摘要:深度学习中使用的训练的数据集中可能存在一些用户的敏感信息，在模型训练过程中，这些敏感信息将会隐含地存在于模型参数中，从而存在隐私泄露的风险。本文在AdamP优化器的基础上引入高斯机制的差分隐私，提出了一种基于一次幂函数的动态隐私预算分配算法来更合理地分配差分隐私的隐私预算，即DP-AdamP，以更好地平衡隐私性和模型准确性。实验结果表明，本文的DP-AdamP算法相比于传统的DP-SGD算法，在相同隐私预算下具有更好的准确性，低隐私预算情况下高出约7.7%，高隐私预算情况下高出约3.9%，并且针对更具有实际意义的低隐私预算的情况进行单独实验，进一步验证了DP-AdamP算法的有效性。 |
| 中文关键词:深度学习 差分隐私 动态隐私预算 AdamP算法 |
| |
| Dynamic Differential Privacy Preserving Algorithm Based on Deep Learning |
|
|
| Abstract:There may be some sensitive information of users in the training dataset used in deep learning, and during the model training process, these sensitive information will implicitly exist in the model parameters, thus there is a risk of privacy leakage. In this paper, we introduce differential privacy of Gaussian mechanism on the basis of AdamP optimizer, and propose a dynamic privacy budget allocation algorithm based on primary power function to allocate the privacy budget of differential privacy more reasonably, i.e. DP-AdamP, so as to better balance the privacy and model accuracy. Experimental results show that the DP-AdamP algorithm in this paper has better accuracy than the traditional DP-SGD algorithm under the same privacy budget, about 7.7% higher in the case of low privacy budget, and about 3.9% higher in the case of high privacy budget, and separate experiments are conducted for the more practically relevant case of low privacy budget, which further validates the effectiveness of the DP-AdamP algorithm. |
| keywords:deep learning differential privacy adaptive privacy budget AdamP algorithm |
| 查看全文 查看/发表评论 下载pdf阅读器 |
|
|
|
