Research On Resource Management And Privacy Preservation For Big Data Driven Mobile Edge Networks

The mobile communication system has recently emerged as essential social infrastructure,providing rich and convenient mobile services covering all aspects of people's daily lives.By decentralizing resources,contents,and network functions at the edge nodes nearing mobile users or data sources,Mobile Edge Computing(MEC)has been widely used to ensure low-latency and high-efficiency service provisioning.MEC provides system support for data collection,content storage,and task processing at the network edge,bringing supplement and innovation to the traditional mobile cloud computing paradigm involving the remote data centers only.The generation and interaction of massive real-time service data and the performance enhancement of intelligent terminal devices further promote the maturity of data-driven network technologies,including edge sensing,edge caching,and edge computing.Besides,numerous Artificial Intelligence(AI)tasks are enabled to be performed in the edge network,which injects more intelligence into the mobile communication system.However,these emerging technologies will bring tremendous pressure to the limited system resources in the mobile edge network,making the multi-dimensional heterogeneous resources such as communication,computing and storage highly coupled and interact with each other.Moreover,to make resource management decisions customizing to the real system relies on massive user data,while the collection and transfer of massive application data will inevitably raise privacy leakage concerns.Motivated by these facts,this thesis focuses on big data-driven mobile edge networks with MEC,especially the four typical mobile systems supported respectively by edge sensing,edge caching,edge computing,and edge intelligence technologies,where we investigate the system resource management and data privacy protection problems toward the specific system requirements.We target reducing operation costs,improving energy efficiency,enhancing robustness,and ensuring security to facilitate the deployment and implementation of the above four network technologies in the new-generation mobile communication system in an efficient,reliable,and ubiquitous way.The main contents and contributions of this thesis are summarized as follows:1.We study how to build a coverage-aware mobile crowdsensing system based on the users' location data while preserving the location privacy of participating users.Based on local differential privacy,we propose a two-stage location privacy protection mechanism against “Bayesian Inference” and “Optimal Inference” theoretically.With our mechanism,each participant can perturb his location data before publishing with a personalized privacy budget without exposing the real location information.Since location perturbation may introduce uncertainty into the collected sensing data,we formulate a robust participant recruitment optimization problem to hedge such uncertainty and maximize the expected overall sensing coverage.Based on the submodular optimization theory,we develop a low-complexity algorithm to determine a robust participant recruitment strategy under a certain recruitment budget.The algorithm has a guaranteed approximation ratio,and the derived strategy can be tolerant to the participants' location shift.Simulation results based on real location datasets verify the efficacy of our location privacy protection mechanism and the participant recruitment scheme in guaranteeing location privacy and improving the decision robustness and the system utility.2.We study how to determine the caching policy using the users' content request records in an edge caching system while preserving users' preferences.With users' random requests,we formulate a robust content caching optimization problem to improve the system energy efficiency and hedge the uncertainty of request preference,where the content popularity with bitrate preference and the constraints of the edge cache capacity are integrated into the optimization problem.Then we leverage the centralized differential privacy technology to protect the privacy of the user's historical request records and characterize the uncertainty of real request arrivals according to the perturbed video request records.By extracting the distributional features of these historical data,we develop a data-driven distributionally robust optimization method to determine the content caching strategy robust to uncertain users' preferences.Simulation results based on real-world datasets show that our method can effectively relieve the risk caused by the uncertainty of content request arrivals while protecting user preferences' privacy.It is also shown that the method exhibits excellent performance gains in improving system energy efficiency and cache hit rate.3.We study how to efficiently deploy computing resources at edge nodes based on historical service data in an edge computing system while preserving the service traces' privacy.Considering the random arrival of service requests,we formulate a risk-averse resource placement problem to improve the revenue of the application service provider and hedge the uncertainty of service demand.Then,we exploit historical service traces at the edge sites to characterize the demand uncertainty in a data-driven manner.To solve the formulated problem without compromising the data privacy,we propose a distributed algorithm that enables each edge site to keep the historical service data locally and participate in the optimization process.The simulation results based on real-world data show that our method can effectively hedge the demand uncertainty under the premise of protecting the privacy of local service data,thus significantly improving the system robustness and the utility of edge service provisioning.4.We study how to train machine learning models over wireless networks based on edge devices' data while preserving the training data privacy.To ease the communication burden during training,we improve the original federated learning algorithm with data isolation and privacy protection characteristics by integrating modern communication compression technologies.We analyze the convergence rate and communication complexity of the proposed training algorithm theoretically.We characterize the impact of compression parameters on the communication/computing load and the energy consumption of edge devices.We then define a compression control optimization problem to minimize the energy consumed by the edge devices for undertaking the training task.By integrating the Benders Decomposition method and the Inner Convex Approximation method,we design a low-complexity compression control algorithm,which gives each participating device a guide on flexibly determining the communication compression parameters.The simulation results validate the correctness of our theoretical analysis and the efficiency of the proposed algorithms.The results also show that the proposed federated training and compression control algorithms have great potentials in adapting to the heterogeneity of participating devices and improving the system energy efficiency.