| Artificial intelligence(AI)technologies are continuously making progress,driven by the law of scale.As neural network models expand in size,with the proliferation of neurons and deeper and broader network hierarchies,the requisite computational power surges exponentially.It is noteworthy that AI is being integrated into various entities of mobile network infrastructure,including radio access networks,core networks,transport networks,and terminal devices.AI is showing a broad range of application prospects in improving access efficiency,expanding network capacity,optimizing channel information feedback mechanisms,precise positioning,and beam management.With the support of EC,the capabilities of AI can be extended to radio access networks and various terminal devices,thereby achieving distributed edge intelligence.However,the complexity of multi-party collaboration not only pose severe challenges to the efficient distribution,training,and deployment of models but also raise a series of issues related to computational security,the trustworthiness of data and model sharing,and privacy protection.Building a trustworthy EC environment is a prerequisite for realizing edge intelligence.Currently,blockchain-based distributed ledger technology(DLT)and consensus mechanisms are considered as key enablers.Through blockchain,the distributed architecture can ensure data integrity,tamper-proofing,and verifiability,making multi-party data exchange and computation processes traceable and auditable.Despite the appeal of blockchain in promoting trusted EC,it still confronts numerous challenges.On the one hand,it is difficult to meet the extreme demands of fast consensus,high-concurrent transactions,low latency,and easy scalability using native blockchain.On the other hand,in multi-party collaboration,model dissemination and deployment must adapt to the distributed nature of data and computing power.Additionally,more concealed model attacks may pose risks to edge intelligence.For the multi-party trust,privacy and efficiency,the main work and conclusions of this dissertation are summarized below.(1)Prior to deploying a blockchain-enabled EC,it is necessary to address the question of how much trustworthiness(probability of successful attacks)can be achieved in a non-trusted environment,as well as the impact of a non-trusted environment on DL performance.To this end,this dissertation first carefully examines the fundamental characteristics and limitations of FL and blockchain in communication networks.The delay incurred in model training,blockchain service,and model transmission between clients significantly affects the free-riding success probability and thus the learning performance.Then,this dissertation analyzes the delays experienced by transactions with and without free-riders based on queuing theory.Further,this dissertation theoretically derives the free-riding success probability and performance difference of global models with and without free-riders,and reduces the impact of free-riders by optimizing the service rate.Finally,theoretical analysis and simulation experiments show that the system suffers considerable performance degradation when subjected to free-rider attack,and the dominant factors affecting FL are identified and studied.(2)Despite blockchain’s ability to ensure the authenticity and credibility of data records,it cannot fully address the issue of interference from untrustworthy participants towards legitimate ones.In fact,operations conducted outside the chain,such as local training and model aggregation,do not rely on direct blockchain records,thus facing trust deficiencies.This dissertation investigates how to achieve trustworthy distributed learning without relying on third-party authorities,aiming to solve the trust mismatch inside and outside the blockchain.To this end,this dissertation comprehensively considers the outcomes of interactions involving trust,distrust,and uncertainty,and designs a probabilistic multi-class trust assessment model.It maps the blockchain-verified interaction results into a ternary trust score based on a multinomial distribution.Considering the impact of insufficient interaction information on trust scores,this dissertation derives a knowledge deficiency coefficient to mitigate this adverse effect.Finally,the dissertation technically validates the effectiveness of the trust mechanism in supervising operations both inside and outside the blockchain,demonstrating superior contribution rates compared to benchmark schemes.(3)Distributed ledger technology and consensus mechanisms facilitate collaboration among edge nodes based on a decentralized trust framework.However,this also introduces challenges of scalability and resource constraints,making edge nodes susceptible to capacity bottlenecks.This dissertation explores the capacity limitations of edge nodes based on sharded blockchain technology and discusses how to design an effective user admission control mechanism.First,this dissertation utilizes the sharding technique to divide the entire edge computing network into multiple smaller logical partitions,thus limiting the model throughput within each partition.Then,this dissertation analyzes the performance indicators such as blocking probability and queuing length under different model aggregation methods,which provides theoretical guidance for improving the system throughput performance.Based on this,this dissertation proposes an admission control strategy based on maximizing model throughput to select edge users with greater contribution to model aggregation and higher efficiency to avoid throughput performance degradation under high concurrency.Simulation results show that the proposed strategy maximizes the model throughput(4.54% ～ 71.47%)compared to the benchmark scheme,and maintains the learning performance(-3% ～ 2.5%).(4)Distributed learning effectively reduces the risk of data privacy leakage in an untrusted environment by sharing trained models among participants rather than sensitive raw data.To address heterogeneity and communication-efficiency issues,this dissertation employs a hierarchical clustering learning(HCL)computing architecture.HCL extends traditional FL by clustering heterogeneous users via cluster nodes(CNs)located at network edge.Unlike one-to-many transmissions,this dissertation proposes a multistage cooperative model dissemination strategy to sequentially determine the subsets of CNs that can concurrently transmit models during individual scheduling stages,thereby improving communication efficiency.The strategy aims to minimize the maximum completion time of the slowest straggler in communication rounds,while accurately clustering users to CNs with similar data distributions.To make sequential and combinatorial decisions in individual stages,this dissertation develops an online learning algorithm to learn the strategy.Furthermore,the strategy dynamically re-clusters users to appropriate CNs,according to the similarity of users’ data distribution.Numerical results demonstrate the superiority of our proposed strategy over some benchmarks in terms of communication(2.11% ～ 5.57%)and learning efficiency.In summary,this dissertation systematically investigates the trust challenges in complex heterogeneous EC environments.Theoretical analysis is conducted on the boundaries of trustworthiness and effectiveness in blockchain-based trustworthy edge intelligence.A series of issues,including access control,model dissemination and model deployment,are optimized in resource-constrained EC environments,thereby improving trust,privacy and efficiency. |
PDF Full Text Request