差分隐私GAN梯度裁剪阈值的自适应选取方法Adaptiveselection.PDFVIP

第 4 卷第 5 期 网络与信息安全学报 Vol.4 No.5 2018 年 5 月 Chinese Journal of Network and Information Security May 2018 差分隐私 GAN 梯度裁剪阈值的自适应选取方法 1,2 1,2 1,2 1,2 郭鹏 ，钟尚平 ，陈开志 ，程航 （1. 福州大学数学与计算机科学学院，福建 福州 350116 ； 2. 网络系统信息安全福建省高校重点实验室，福建 福州 350116 ） 摘 要：提出自适应选取差分隐私 GAN 梯度裁剪阈值的方法。该方法假设可以接触到与隐私数据同分布的 小部分公开数据，通过从公开数据中随机选取一批数据，设置裁剪阈值为这批数据的平均梯度范数，迭代上 述操作直到网络聚合。在 Mnist 和 Cifar10 数据集上对所提方法进行了实验验证，结果表明，在合理隐私预 算下与差分隐私辅助分类 GAN 相比，卷积神经网络（CNN ）分类器准确率提高 1%~4%，而评估分数（inception scores ）提升0.6~1.2 。 关键词：生成对抗网络；差分隐私保护；梯度裁剪阈值；自适应选取 中图分类号：TP309 文献标识码：A doi: 10.11959/j.issn.2096-109x.2018041 Adaptive selection method of differential privacy GAN gradient clipping thresholds GUO Peng1,2, ZHONG Shangping1,2, CHEN Kaizhi1,2, CHENG Hang1,2 1. College of Mathematics and Computer Science, Fuzhou University, Fuzhou 350116, China 2. Network System Information Security Fujian Provincial University Key Laboratory, Fuzhou 350116, China Abstract: A method of adaptive selection of differential privacy GAN gradient clipping threshold was proposed. The method assumes that a small portion of public data that is identically distributed with the private data can be accessed, a batch of data is randomly selected from the public data, a clipping threshold is set as an average gradient norm of the batch of data, and the above operations are iterated until the network converges. The method was veri- fied on the Mnist and Cifar10 data sets. The results show that under a reasonable privacy budget, the accuracy of CNN classifiers is improved