近年来，随着手机、平板等高性能移动设备的飞速普及和发展，全球社交网络用户逐年呈爆炸式增长。2022年全球社交网络用户有37.5亿人，这意味着全球近二分之一的人在使用社交网络平台的服务。由于工作或社交需求，社交网络用户频繁发布和分享大量信息，因此产生了海量的社交数据。这些数据规模庞大并且相互交错重叠，给社交网络服务平台的访问控制带来了巨大的挑战。为了应对这个问题，亟需一种高效性的访问控制模型来提高社交网络平台的数据访问控制能力，减轻平台数据压力，为用户提供良好的交互体验。为了提高社交网络平台的访问控制效率，本文具体的研究内容如下：

通过研究分析社交网络访问控制策略集规模庞大和复杂度高的特征，本文建立了集中式访问控制模型DPEngine。首先，DPEngine中提出了一种能够适应社交网络策略集特性的高效聚类算法。其次，DPEngine通过该聚类算法对策略集的处理，采用策略索引和相似度计算的方法来减少策略评估的规则比较次数，从而加快了社交网络平台的策略评估速度，提高了社交网络平台的数据访问控制能力。实验结果表明，与现有的访问控制模型Sun PDP、HPEngine、XEngine和SBA-XACML相比，DPEngine的评估效率分别提高了90.23%、80.47%、63.67%和74.06%。随着请求数量的增加，DPEngine的评估时间呈线性增长，因此DPEngine具有高效率和稳定性的优势。

由于集中式访问控制模型的评估效率随着策略集规模增大容易遇到瓶颈，采用分布式机器学习的思想，引入了自适应轮循分组和梯度合并的异步策略，建立了分布式社交网络访问控制模型DSEngine。首先，DSEngine通过轮循分组的异步策略将计算与通信时间重叠，减少多节点并行计算网络争用导致的资源浪费，加快多节点协同进行策略评估的速度。其次，DSEngine采用梯度合并的方法来减少节点的梯度传输时间，加快了分布式模型的匹配计算速度，从而提高了社交网络访问控制平台的策略评估性能。实验结果表明，本文提出的分布式访问控制模型DSEngine的评估效率比现有的分布式模型DeepSpeed-ZeRO，Megatron-LM，Paddle与LP-SBA-XACML分别提高了22.3%、26.47%、33.6%和24.1%，表明了DSEngine在面对大规模社交网络访问控制策略评估问题上的高效性。

本文在经典的访问控制模型基础上，结合实际社交网络中用户和数据量规模庞大的特点，建立了适应社交网络平台大数据环境下的访问控制模型，有效提高了社交网络平台的访问控制效率，为高效社交网络访问控制系统的研究提供了借鉴。

In recent years, with the rapid development of high-performance mobile devices, global social network users have shown an explosive growth. In 2022, there were 3.75 billion global users, which means that nearly half of the world’s population is using social network platform services. Due to the work or social requirements, social network users frequently post and share a large amount of information, thus producing a variety of social data. These massive information are huge and overlapping, resulting in significant challenges to the access control of social network service platforms. In order to address the above issues, an efficient access control model is urgently anticipated to improve the access control capabilities, which will reduce the data pressure and provide excellent interaction experiences for social network users. In order to improve the access control efficiency of social network platforms, the specific research contents of this thesis is as follows:

1) By analyzing the characteristics of the large scale and high complexity of social network access control policy sets, this thesis establishes a centralized access control model DPEngine. First, DPEngine proposes an efficient clustering algorithm which can adapt to the characteristics of social network policy sets. Second, by using the cluster method to process the policy sets, DPEngine uses policy indexing and similarity calculation methods to reduce the number of rule comparisons in policy evaluation, thereby accelerating the speed of policy evaluation on social network platforms and improving access control capabilities. The experimental results show that compared with the existing access control models, such as Sun PDP, HPEngine, XEngine, and SBA-XACML, the evaluation efficiency of DPEngine has increased by 90.23%, 80.47%, 63.67%, and 74.06%, respectively. With an increasing number of requests, the evaluation time of DPEngine increases linearly, thus DPEngine has the advantages of high efficiency and stability.

2) Due to the situation that the evaluation efficiency of centralized access control models is easily bottlenecked with the increase in policy set size, this thesis introduces the idea of distributed machine learning, asynchronous strategies of adaptive round grouping and gradient merging to establish a distributed social network access control model DSEngine. First, DSEngine utilizes an asynchronous strategy of round grouping to overlap computation and communication time, reducing resource waste caused by multi node parallel computing network contention, and accelerating the speed of multi node collaborative policy evaluation. Second, DSEngine adopts gradient merging method to reduce the gradient transmission time, accelerate the rule matching calculation speed of distributed models, thereby improving the policy evaluation performance of social network access control platforms. The experimental results show that the policy evaluation efficiency of the distributed access control model DSEngine proposed in this thesis is 22.3%, 26.47%, 33.6%, and 24.1% higher than the existing distributed models DeepSpeed-ZerO, Megatron-LM, Paddle, and LP-SBA-XACML, respectively. This indicates a high efficiency of DSEngine in evaluating access control policies in large-scale social networks.

This thesis establishes an access control model adapted to the big data environment of social network platforms based on classic access control models, it effectively improve the access control efficiency of social network platforms, and provide a reference for the research of efficient social network access control system.