Research On Some Key Problems Of Edge Computing Based Mobile Crowd Sensing Systems

With the development of wireless technology and popularization of mobile smart devices,mobile crowd sensing system has become a new sensing paradigm in the Internet of Things.In such systems,the ubiquitous smart mobile devices,such as the smart phones,wearable devices,smart vehicles,etc.,which are embedded with numerous sensors,such as the cameras and gyroscopes,are applied to collect a variety of sensory data.At present,mobile crowd sensing has become a very attractive paradigm to collect and analyse the large-scale data in a crowded area.It has become an indispensable component of modern smart city due to its powerful and effective capability of data collection.In addition,the increasing numbers of smart devices and their embedded sensors have promoted the mobile crowd sensing to be rapidly developed in recent years and widely applied to other fields,such as the financial services,traffic management,healthcare and so on.The increase in the number of mobile devices allows the mobile crowd sensing to be applied to more and more complicated sensing tasks that need a volume of sensory data.However,transmitting and analysing such large-scale data may introduce a serious network delay,which results in unsatisfactory experience in some low-latency services.In order to improve the efficiency of mobile crowd sensing,an edge computing based mobile crowd sensing system has been proposed and attracted more and more attention.Such new systems utilize the capabilities of storage,computation,and transmission of edge nodes deployed between the mobile devices and central cloud-based platform to greatly reduce the computing and storage burden of platform and mobile devices.At the same time,due to the existence of edge nodes,the systems successfully reduce the transmission delay and network congestion,while greatly improving the efficiency of data processing and analysis.Although edge computing based mobile crowd sensing systems have been applied to many practical applications,they still have many important problems that need to be solved.The first problem is to explore the computing capability of smart devices: With the development of communication technology,smart devices are becoming more and more plentiful.At present,many smart mobile devices have embedded with the computing engines,such as the Neural Engine in i Phone and the Neural Processing Engine in Snapdragon.However,the traditional mobile crowd sensing systems are usually applied to simple data collection,which can not fully explore their computing and processing capabilities.Therefore,it requires us to employ the mobile crowd sensing to other applications that require higher computing capabilities,such as the model training in federated learning.The second is to design an incentive mechanism: Applying the mobile crowd sensing systems to the scenarios that require more powerful computing capability mentioned in the first problem or the scenarios for massive data collection needs to consider the incentive requirement since the systems rely on a large number of workers voluntarily participating in the execution of sensing task.Therefore,the incentive mechanism needs to be designed to attract more participation to improve the quality of collected data.An appropriate mechanism also guarantees the efficiency and stabilization of systems.The third problem is to design a data storage mechanism: Although as mentioned in the first problem,we aim to apply the mobile crowd sensing systems to some scenarios that require more powerful computing capability,the systems are presently applied to the traditional data collection.However,with the increasing number of workers and data requesters in mobile crowd sensing systems,the volume of data that needs to be processed and transmitted by the platform is becoming larger and larger.Furthermore,the massive amount of data sent to the requesters usually causes a serious network congestion and transmission delay.Therefore,to solve these problems,it is necessary to design a data caching mechanism to store some popular data at the edge nodes such that the data requesters can obtain the requested data from these edge nodes without communicating with the platform.This thesis has conducted a more comprehensive research focusing on the above three issues in the edge computing based mobile crowd sensing systems.Our main works are as follows:First,in order to explore the powerful computing capability of mobile smart devices,a novel framework of federated learning in mobile crowd sensing systems is proposed,namely,SHIELD,by jointly considering the incentive requirement,privacy protection,and high training accuracy.In fact,SHIELD bears the properties of truthfulness and individual rationality such that it is able to attract more participants.Furthermore,it meets the differential privacy of workers’ reported costs.Additionally,SHIELD also guarantees the differential privacy of local training models.As to the training accuracy,the excess empirical risk of SHIELD is tightly upper bounded,where a special case for the totally distributed scenarios leads to a much sharper bound than the latest known result.Finally,numerous experiments in classification and regression tasks show that comparing with the state-of-the-art approaches,the training accuracy of SHIELD is increased by approximately 80% at most in some cases.Second,since the participation incentive plays an indispensable role in our first contribution and other mobile crowd sensing applications,we further focus on the incentive mechanism design in edge computing based mobile crowd sensing system.However,in such systems,all data requesters,workers and edge nodes behave strategically and selfishly to maximize their own utility.Furthermore,since the platform and workers communicate with each other through the edge nodes,their transmitted information may be manipulated by the edge nodes.Therefore,to solve these problems and attract more participants to join in the systems,we propose a novel incentive mechanism,namely,CHASTE.In fact,CHASTE ensures that all requesters,workers,and edge nodes behave truthfully.Furthermore,CHASTE allows them to receive a non-negative utility once they participate in the systems,i.e.,it satisfies the individual rationality.Apart from the truthfulness and individual rationality,CHASTE also guarantees the budget balance and achieves a high social welfare.Finally,extensive simulations are conducted to validate the desirable properties of CHASTE,where the results show that comparing with the state-of-theart approaches,the social welfare achieved by CHASTE is increased by approximately 23.8%in some cases.Third,since the most mobile crowd sensing systems are presently applied to the massive data collection,the requesters require to communicate with the platform frequently to obtain the huge amount of requested data,which may lead to heavy network congestion and delay.To tackle this problem,some small base stations with caching capability are employed in the edge computing based mobile crowd sensing systems to store some popular data,so that the data requesters can obtain the requested data from the nearby small base stations.As the underlying infrastructure of mobile crowd sensing systems,wireless networks have two characteristics called the requester mobility and wireless interference.Therefore,by considering the impacts caused by these two characteristics together,we propose two data caching mechanisms,namely,the short length coded caching and long length coded caching.The explicit expressions of data offloading ratio are derived for the short length coded caching and long length coded caching in the environments with requester mobility and wireless interference.Furthermore,the maximization of data offloading ratio is then investigated and a sub-optimal solution is obtained in this NP-hard problem.Additionally,the corresponding data offloading ratio in a degraded environment that only considers the requester mobility is further investigated,where a suboptimal solution with linear computational complexity is obtained utilizing the submodularity property of their expressions.Finally,some extensive simulations are conducted to validate the desirable properties of our proposed caching mechanisms.The simulation results show that comparing with the state-of-the-art approaches,the data offloading ratios of long length coded caching and short length coded caching are averagely increased by 22.7% and 10% in some cases,respectively.