【作者】 刘翔；

【导师】 郭宇春；

【作者基本信息】 北京交通大学 ， 通信与信息系统， 2022， 博士

【摘要】 移动边缘网络(Mobile Edge Network)作为一种新兴的网络架构,正受到学术界和工业界的广泛关注,各种研究和系统层出不穷。它的核心思想是将计算和存储资源部署在更靠近用户的网络边缘,并利用软件定义网络、网络功能虚拟化等技术对各项资源进行灵活分配与扩展,实现按照用户需求优化分配资源的目标。移动边缘网络的发展目前遇到两方面的挑战:(1)用户的移动导致用户需求的时空分布高度动态,这不仅增加了需求感知的难度,而且对资源分配策略的高效性和实时性提出了更高要求;(2)对用户需求的测量感知会带来隐私泄露风险。为了应对这些挑战,本文从移动边缘网络的实际需求出发,基于海量用户轨迹大数据,从移动模式挖掘、用户需求感知、资源协同优化、用户隐私保护等方面进行了深入研究,提出了解决方案,并实验验证它们对提升网络服务质量、降低网络资源开销、保证用户隐私安全方面的有效性。具体工作如下:(1)基于轨迹表征的群体移动模式挖掘为刻画和挖掘用户需求的时空关联特征,本文提出了一种基于轨迹深度表征的群体移动模式挖掘方法,在海量真实轨迹大数据中实现了准确的轨迹相似度计算和高效的热门出行线路挖掘。具体来说,1)针对现有轨迹表征模型存在轨迹表征不准确、不全面的问题,提出了一种基于对比自监督的轨迹表征模型CSTRM(Contrastive Self-Supervised Trajectory Representation Model)。该模型有两个特点:a)设计了基于对比学习的损失函数,能够全面捕获轨迹中轨迹级别特征和点级别特征;b)提出了基于自监督的数据增广方法,结合下采样、轨迹扰动和负采样的手段,实现了训练数据的高效生成,有效解决了训练数据匮乏的问题。在真实用户轨迹大数据上的实验结果表明:CSTRM在三种轨迹相似度计算任务上,比现有模型获得了最高15%的准确度提升,并且轨迹表征向量的分布范围更广。2)基于CSTRM获得的轨迹深度表征,设计了基于轨迹聚类的群体移动模式挖掘算法,发现了城市中的热门出行线路,为设计资源优化策略奠定了基础。(2)用户需求感知的资源协同优化算法针对用户高速移动的移动边缘网络场景中,用户需求与资源部署存在失配的问题,本文设计了用户需求感知的资源协同优化算法。该算法包括两个部分:1)用户需求预测模型。设计了基于多步预测的用户移动模型和基于观看行为的用户请求模型,以准确预测一个缓存更新周期内因用户多次位移产生的用户需求变化;2)资源协同优化算法。考虑用户需求的时空分布,在基带单元计算资源和网络缓存空间均受限的条件下,实现基带和射频单元的动态匹配策略和内容部署策略的协同优化,最大化系统缓存命中率。为此,本文首先将该问题建模为一个整数线性规划问题,证明了其NP-hard特性,然后提出基于热门线路的静态优化策略和基于贪婪合并的动态优化策略,高效求解该优化问题。在真实数据上的仿真结果表明:1)相比于基于内容流行度和基于用户兴趣的用户需求预测模型,本文提出的用户需求预测模型能更准确刻画用户需求的时空分布,使得缓存优化策略分别获得至少48.1%和51.1%的性能提升。2)本文提出的静态和动态优化策略分别比基准优化策略获得至少7%和2.6%的缓存命中率提升。(3)时空数据流中用户轨迹隐私保护算法为了准确感知用户需求的时空动态分布,移动边缘网络不断统计聚合用户移动轨迹和浏览轨迹,构建区域热度流和内容热度流。为了防止攻击者从这些聚合的时空数据流中推测出用户的轨迹,保证用户隐私安全,本文设计了轨迹隐私保护算法。针对现有隐私保护模型忽略空间维度,无法兼顾用户轨迹隐私保护和聚合数据可用性(准确性)的问题,本文提出基于二维滑动窗的轨迹隐私保护框架。该框架有三个特点:1)时空二维窗口。考虑用户移动轨迹和浏览轨迹的时空连续性,定义了结合n-range空间窗口和w-slot时间窗口的时空二维窗口(w,n);2)基于二维滑动窗的(w,n)-差分隐私保护框架。其中包括两项独特设计:a)通过扰动时空窗口内的数据,保证攻击者无法准确获知其中任意一条时空序列的信息;b)在时空数据流中不断滑动时空窗口,确保任意一条处在连续w个时间窗口以及n-range空间范围内的轨迹序列的隐私安全。3)两种(w,n)-差分隐私保护算法。两种算法通过感知数据流的时空动态变化特征,制定动态的数据扰动与发布策略,提升了数据可用性。理论分析表明,两种算法均满足(w,n)-差分隐私保护框架的要求。在真实用户轨迹大数据上的实验结果表明,两种算法在保证轨迹隐私安全的基础上,分别获得了至少18%和30%的数据可用性提升。更多还原

【Abstract】 As an emerging network architecture,Mobile Edge Network(MEN)is receiving extensive attention from academia and industry.Its core idea is to deploy computing and storage resources at the edge of the network closer to users,and use software-defined networking,and network function virtualization to flexibly allocate and expand resources,to realize the adaptation of network resources according to user demands.The development of MEN currently encounters two challenges:(1)The movement of users leads to a highly dynamic spatial-temporal distribution of user demands,which not only increases the difficulty of demand perception,but also puts forward higher requirements on the efficiency and real-time performance of resource adaption.(2)The perception of user demands will bring the risk of privacy leakage.To satisfy the actual needs of mobile edge networks and solve the above challenges,this thesis conducts in-depth research on mobility pattern mining,user demand perception,resource collaborative optimization,and user privacy protection,and provides effective solutions to achieve the purpose of improving network service quality,reducing network resource overhead,and ensuring user privacy and security.The specific contributions are summarized as follows:(1)Group Mobility Pattern Mining Based on Trajectory RepresentationTo find the spatio-temporal correlations of user demands,this thesis proposes a group mobility pattern mining method based on trajectory representation,which realizes accurate trajectory similarity computation and efficient popular routes mining in massive trajectory datasets.Specifically,1)Aiming at the problem of inaccurate and incomplete trajectory representation in existing trajectory representation models,this thesis proposes a novel Contrastive Self-Supervised Trajectory Representation Model.This model has two characteristics: a)a contrastive loss function is designed to capture both trajectorylevel features and point-level features in the trajectories;b)a self-supervised data augmentation method is proposed.By combining dropping,distorting,and negative sampling,this method can efficiently generate training data The experimental results on real trajectory datasets show that CSTRM achieves a maximum accuracy improvement of 15% over existing models in three trajectory similarity computation tasks,and the distribution of trajectory representation vectors is wider than existing models.2)Based on the accurate representations obtained by CSTRM,a clustering-based group mobility pattern mining algorithm is proposed,which can efficiently discover the popular routes in the city.(2)Demand-aware resource collaborative optimization algorithmAiming at the problem of mismatch between user demands and resource deployment in the mobile edge network scenario where users move at high speed,this thesis designs a demand-aware resource collaborative optimization algorithm.This algorithm consists of two parts: 1)User demand prediction model.To accurately predict the change in user demand caused by multiple movements of users in one cache update cycle,this thesis proposes a multi-step user mobility prediction model and a behavior-based user request prediction model.2)Collaborative optimization algorithm.Considering the spatialtemporal distribution of user demands,the matching strategy and content deployment strategy of the BBUs and RRHs are studied to maximize the system cache hit rate under the constraints of BBU utility and network cache space.To this end,this thesis first formulates the problem as an integer linear programming problem and proves its NP-hard property.Then,a popular routes-based static optimization strategy and a greedy mergebased dynamic optimization strategy are proposed to efficiently solve this optimization problem.The simulation results on real data show that: 1)Compared with popularitybased and preference-based models,the proposed user demand model can accurately predict the spatial-temporal distribution of user demand and obtain 48.1% and 51.1%improvement on cache hit rate.2)The proposed static and dynamic optimization strategies achieve at least 7% and 2.6% cache hit rate improvement over the optimal baseline strategy,respectively.(3)Trajectory Privacy Protection Algorithm in Spatio-temporal Data StreamTo accurately sense the spatio-temporal distribution of user demands,the mobile edge network continuously aggregates user data and counts the number of users in each region and the number of visits to each content.To prevent attackers from inferring the user’s trajectory from these aggregated spatio-temporal data streams,this thesis designs a trajectory privacy protection framework.Aiming at the problem that the existing privacy protection models ignore the spatial dimension and cannot strike a balance between trajectory privacy protection and stream data utility(accuracy),this thesis proposes a trajectory privacy protection framework based on a two-dimensional sliding window method.This framework has three characteristics: 1)spatial-temporal window.Considering the spatio-temporal continuity of user trajectory,a spatio-temporal window(w,n)is defined by combining the n-range spatial window and the w-slot temporal window;2)(w,n)-difference privacy framework based on sliding window methodology.This framework includes two unique designs: a)By perturbing the data in the spatiotemporal window,it ensures that the attacker cannot infer any spatio-temporal sequences within this window;b)By sliding the spatio-temporal window on the data stream,this model ensures the safety of any trajectory sequences within w consecutive timestamps and n-range places.3)two(w,n)-differential privacy algorithms.To improve data utility,two algorithms detect the spatial-temporal dynamics of the data stream and adaptively perturb and publish the data stream.Theoretical analysis shows that two algorithms satisfy the(w,n)-differential privacy.The experimental results show that the algorithms have achieved at least 18% and 30% improvement in data utility while ensuring trajectory privacy.更多还原