摘要
为防止加权社交网络恶意攻击侵犯,提出一种深度差分隐私数据保护算法。首先通过卷积层得到多个特征图,利用池化层将特征图降维,加强深度学习模型处理高维数据的能力,减少网络系统能耗;其次将深度学习模型与差分隐私原理融合,使用差分隐私和高斯分布可组合特点,在运算时结合差分隐私理论,同时加入噪声机制,生成深度差分隐私数据算法;最后通过分析初始GAN的训练过程,得到一个满足维纳-霍普夫方程原理的生成器模型,以此得到攻击者可能获取的最佳结果伪造样本,对深度差分隐私模型实施参变量调优,实现数据集的可用性和隐私保护的均衡性。仿真结果表明,生成的图像辨识度下降,隐私保护力度呈现上升趋势,噪音绝对误差和平均查询绝对误差均减小,实现了深层次保护隐私数据安全的目的,具有较高的实用价值。
In order to prevent malicious attacks on weighted social network, an algorithm of deep differential privacy data protection was put forward. Firstly, multiple feature maps were obtained by convolution layer. Secondly, the pooling layer was used to reduce the dimension of feature map and strengthen the ability of deep learning model to process high-dimensional data, and thus to reduce the energy consumption of network system. Thirdly, the deep learning model was integrated with the principle of differential privacy. Because the differential privacy and Gaussian distribution were composable, the theory of differential privacy was combined with the noise mechanism to generate the deep difference privacy data algorithm. Finally, a model of generator meeting the principle of Wiener-Hopf equation was built by analyzing the training process of Gan Initialization, so that the best forged samples that the attacker might get could be obtained. In addition, the parameters of deep differential privacy model were optimized to achieve the availability of data set and the balance of privacy protection. Simulation results show that the recognition degree of the generated image is decreased. The intensity of privacy protection is rising. The absolute error of noise and the absolute error of average query are reduced. Thus, the purpose of deep protection of privacy data security was achieved. This algorithm has high practical value.
作者
周硙
ZHOU Ai(Beijing University of Technology,Beijing 100124,China)
出处
《计算机仿真》
北大核心
2020年第10期282-285,373,共5页
Computer Simulation
关键词
加权社交网络
深度差分隐私
深度学习
训练数据集
Weighted social network
Deep differential privacy
Deep learning
Training data set