目录
摘要Ⅲ
关键词Ⅲ
AbstractⅣ
引言
引言 1
1 文献综述 1
1.1 安全多方计算 1
1.1.1 同态加密 2
1.1.2 混淆电路 2
1.2 深度学习 3
1.2.1 全连接层 4
1.2.2 激活层 4
1.2.3 卷积层 4
1.2.4 池化层 5
1.2.5 批归一化层 5
1.3 网络权值量化 6
1.3.1 二值神经网络 6
1.3.2 XNORNet神经网络 8
2 方案论证 8
2.1 需求分析 8
2.1.1 模型训练 8
2.1.2 模型预测 9
2.2 整体结构 9
2.3 程序实现 10
2.3.1 模型训练 10
2.3.2 参数导出 10
2.3.3 不经意推断 10
2.4 高级API 11
3 实验分析 12
3.1 实验准备 12
3.1.1 实验环境 12
3.1.2 数据集介绍 12
3.1.3 网络结构介绍 12
3.2 实验过程 13
3.3 实验结果 14
3.4 分析讨论 15
4 结论总结 15
参考文献16
致谢17
基于混淆电路的不经意推断深度学习框架
摘 要
近几年，深度学习(Deep Learning, DL)在图像、音频的处理与分类等问题上表现优秀，应用广泛。以图像分类为例，基于机器学习即服务(Machine Learning as a Service, MLaaS)模式，用户通过服务提供方提供的API将需要分类的图像数据提交给服务端，服务端使用已经训练好的图像分类深度神经网络模型对接收到的图像进行推断，将推断结果返回给用户。然而，很多时候用户输入的数据包含用户隐私信息，直接暴露给服务端的话存在用 *51今日免费论文网|www.51jrft.com +Q: ^351916072^ 
户个人信息泄漏的安全隐患。为解决该问题，一种被称作不经意推断(Oblivious Inference)的概念被提出。不经意推断允许用户在不将输入数据暴露给服务器的前提上获得神经网络的推断结果，以此保护用户隐私。本文基于混淆电路，实现了一种基于XNORNet的不经意推断深度学习框架。得益于XNORNet将神经网络的权值二值化、使用XNOR和popcount运算代替矩阵乘法加法的优势，本框架在尽可能保证网络推断精度的同时提高了运算速度，与标准全精度神经网络相比，内存消耗降低了85%，运算速度提高了20倍。此外，本框架还实现了一套高级API，方便开发者使用简单的语法快速构建基于本框架的神经网络程序。
OBLIVIOUS INFERENCE DEEP LEARNING FRAMEWORK BASED ON GARBLED CIRCUIT
ABSTRACT
In recent years, Deep Learning (DL) has performed well in image and audio processing and classification, thus has been widely used. Taking image classification as an example, based on the Machine Learning as a Service (MLaaS) mode, users submit image data to be classified to the server through the API provided by the service provider, and the server uses the trained image classification deep neural network model to process the received image and returns the inference result to the user. However, in many cases, the data submitted by the user contains user privacy information, and if that information are exposed directly to the server, there is a security risk of leakage of the users personal information. To solve this problem, a concept called Oblivious Inference was proposed. In Oblivious Inference process, the user obtains the final inference result of the neural network without exposing the input data to the server, thereby protecting the user privacy. In this paper, based on the Garbled Circuit, a stateoftheart deep learning framework based on XNORNet is implemented. Taking advantages of XNORNet, including binarization of the weights of the neural network and using XNOR and popcount operations instead of matrix multiplication and addition, this framework improves the operation speed while properly ensuring the network accuracy. Compared with oblivious inference deep learning framework using standard fullprecision neural networks, the memory consumption is reduced by 85%, and the operation speed is increased by 20 times. This paper also implements a set of highlevel APIs for developers to use simple syntax to quickly create neural network codes using this framework.

原文链接：http://www.jxszl.com/jsj/wljs/606906.html