【作者】 王勇；

【导师】 李瑞轩；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2022， 硕士

【摘要】 随着人工智能技术的发展,当前越来越多的应用试图为移动设备提供智能服务,然而移动设备的便携属性导致其计算资源与电量难以支撑大型深度神经网络(Deep Neural Network,DNN)的运行,当前主流的解决方案是将DNN直接部署在云端服务器,或者将DNN分割成两部分,分别部署在本地移动端和云端进行协同推理,依靠云端服务器充足的计算资源分担本地计算的压力,然而这种方式会将用户的原始数据或未经处理的中间数据暴露给云端,给用户的隐私安全带来极大的隐患。为了解决资源受限的移动终端无法在合理的延迟与能耗情况下部署运行大型DNN模型实现安全高效的推理问题,提出了边缘计算中DNN协同推理的隐私保护方法,主要特点为:(1)采用非侵入式的隐私保护方案,通过自注意力机制对本地端产生的中间结果动态地注入关于输入特征相关的噪声,并基于互信息的损失函数最大限度地抹除与推理任务无关的信息,实现对用户隐私信息的保护,无需对DNN模型进行更改和重新训练;(2)采用基于面向隐私保护的DNN协同推理自适应分区方法,根据运行环境与用户需求动态地选择最佳分区点,构建低延迟、低能耗、高准确率且具有隐私保护功能的协同推理系统。通过在LeNet、AlexNet、CIFAR-10、VGG-16上测试,与基线方法相比,平均降低了34.08%的总体推理时延,减少了移动端17.96%的能耗,通过隐私保护算法将原始输入和传输到云端之间的实际信息减少了74.29%,同时只牺牲了1.57%的准确率损失。对比实验表明,提出的方法在降低延迟与终端能耗上,性能与当前最好的算法相当,但是在隐私保护效果上提升了1.2～3倍。更多还原

【Abstract】 With the development of artificial intelligence technology,more and more applications are trying to provide intelligent services for mobile devices,however,the portable property of mobile devices makes it difficult to support the operation of large deep neural network(DNN)with its computing resources and power.The current mainstream solution is to deploy DNNs directly on cloud servers,or split DNNs into two parts and deploy them on local mobile and cloud for collaborative inference,relying on sufficient computing resources of cloud servers to share the pressure of local computing,however,this approach will expose users’ original data or unprocessed intermediate data to the cloud,which brings great potential risks to users’ privacy and security.In order to solve the problem that resource-constrained mobile terminals cannot deploy and run large DNN models with reasonable latency and energy consumption to achieve secure and efficient inference,a privacy-preserving approach to DNN collaborative inference in edge computing is proposed,with the following main features:(1)A nonintrusive privacy-preserving scheme is adopted,which dynamically injects noise about the input features into the intermediate results generated at the local end through a self-attentive mechanism and maximizes the erasure of information irrelevant to the inference task based on the loss function of mutual information to achieve the protection of user privacy information without changing and retraining the DNN model;(2)The privacy-preserving DNN collaborative inference adaptive partitioning algorithm is used to dynamically select the optimal partitioning points according to the operating environment and user requirements,and to build a collaborative inference system with low latency,low energy consumption,high accuracy and privacy-preserving functions.By testing on LeNet,AlexNet,CIFAR-10,and VGG-16,the overall inference latency is reduced by 34.08% and mobile energy consumption by 17.96% on average compared with the baseline method,and the actual information between the original input and the transmission to the cloud is reduced by 74.29% by the privacy-preserving algorithm,while sacrificing only 1.57% of accuracy loss.Comparative experiments show that the performance of the proposed approach is comparable to the best current algorithms in terms of latency reduction and endpoint energy consumption,but with a 1.2-3 times improvement in privacy-preserving effects.更多还原