Research On Optimization Mechanisms For Federated Learning In Internet Of Things

In the digital era,the artificial intelligence(AI)-enabled Internet of Things(Io T)has become an important component of the modern infrastructure.In the traditional AI training paradigm,massive amounts of data collected by Io T devices need be transmitted to one data centre for centralised AI model training.However,Io T users are often reluctant to share their local data due to data security and privacy concerns.In addition,relevant data security protection regulations have been enacted,which makes it hard to conduct traditional centralised AI model training.Federated Learning(FL)is recognised as an extraordinary solution to overcome this dilemma.FL,an emerging paradigm for AI distributed training,is able to orchestrate the distributed training of AI models across massive Io T nodes while ensuring that users’ local data are not exported to the outside world,and FL has been widely applied in healthcare and industry,etc.However,due to the heterogeneity and limited energy resources of Io T devices,as well as the limited communication resources,FL systems suffer from the issues of low training accuracy,high training delay,high training resource cost,and insufficient incentives for user training.Existing optimization schemes for FL systems in Io T mainly focus on optimizing a single metric,such as training accuracy or training latency,while ignoring concurrent issues such as high training resource cost and insufficient user incentives.In this context,extensively considering the accuracy,latency,resource cost and user incentives of FL distributed training,this thesis focuses on the design of optimization mechanisms for FL systems in Io T.This thesis consists of four researches,which are summarized below.1.Considering the heterogeneity of Io T devices and limited communication resources,this thesis proposes a distributed training optimization mechanism for FL systems in edge networks.In this mechanism,according to the synchronization of FL distributed training,this thesis firstly adopts the idea of clustering to design a heterogeneityaware device scheduling strategy based on an iterative self-organizing data analysis techniques algorithm.This device scheduling strategy can cope well with the problem of device heterogeneity,thereby improving the training accuracy and reducing the training latency for achieving the desired accuracy.Then,by analyzing the property of the optimal communication resource allocation solution,this thesis follows the principle of FL distributed training synchronization and designs a mixed integer linear programming-enabled communication resource management technique to accommodate the limited communication resources,thereby further improving the distributed training latency.Finally,this thesis conducts extensive experiments to verify the the effectiveness of this distributed training optimization mechanism for FL systems in edge networks.2.Further considering the FL performance bottlenecks in edge networks and device energy cost constraints,this thesis proposes a distributed training optimization mechanism for hierarchical FL systems in the Io T-edge-cloud architecture.In this mechanism,this thesis firstly extends the traditional FL system in an edge network to a hierarchical FL system in an Io T-edge-cloud architecture,and constructs the corresponding theoretical foundation to ensure distributed training performance and convergence.Then,this thesis designs a utility-driven heterogeneity-aware heuristic device scheduling strategy to more effectively address the issue of device heterogeneity.Afterwards,by leveraging the slack time in distributed training,this thesis then designs a device operation frequency control technique to optimize the energy consumption cost of participating devices without derating the training accuracy and latency.Finally,this thesis conducts extensive experiments to verify the the effectiveness of this distributed training optimization mechanism for hierarchical FL systems.3.Taking into account the demotivation of user participation caused by resource consumption,this thesis proposes a two-stage energy-aware incentive mechanism for hierarchical FL systems to improve user satisfaction,thereby incentivizing users to actively participate in training.In this mechanism,considering training contributions,this thesis first constructs an energy cost-based user satisfaction model and clarifies the rationality of this model to represent participation motivation.Subsequently,this thesis designs a two-stage energy cost-aware reward allocation strategy,which consists of cloud server and edge server levels.At the cloud server level,based on the amount of energy consumed by all sub-FL systems,the rewards are allocated to the edge servers in proportion to their energy consumption.At the edge server level,based on the heterogeneity and reward preference of Io T users,this thesis designs a water injectiondriven reward allocation technique that optimally distributes rewards to participating users.Finally,this thesis conducts extensive experiments to verify the the effectiveness of this incentive mechanism for hierarchical FL systems.4.To sufficiently eliminate the energy resource constraints of devices and the energy cost concerns of users,this thesis proposes a distributed training optimization mechanism for FL systems with renewable energy supply.In this mechanism,this thesis first analyzes the underlying theory of FL distributed training,and adopts it as the guide to design a multi-armed bandit-based heuristic device scheduling strategy,and then clarifies the rationality of its state,action,reward,and experience feedback strategy,so as to preferentially schedule devices with high contribution to the distributed training.Subsequently,this thesis employs a mixed-integer linear programming-based communication resource management scheme to further optimize the communication efficiency of distributed training.Finally,this thesis conducts extensive experiments to verify the the effectiveness of this distributed training optimization mechanism for FL systems with renewable energy supply.