Research On Federated Split Learning Mechanism For Resource-Constrained IoT Terminals

Distributed learning is a key technique for large-scale machine learning as it enables the training of complex AI models by splitting the task into multiple nodes for parallel execution.The current mainstream distributed learning techniques are federated learning and partitioned learning,which achieve efficient distributed training of multiple terminals-servers through model parameter aggregation and model splitting,respectively.With IoT expected to be the main source of network data in the future,this dissertation explores the development of a federated partition learning mechanism for resource-constrained IoT terminals to enable efficient and safe data training.However,existing federated learning and segmentation learning techniques face challenges related to network resource bottlenecks and data security threats,which cannot meet the model training requirements of resource-constrained massive IoT terminals.Firstly,resource-constrained IoT terminals are unable to support complex neural network model training,and the resource bottleneck of edge servers prevents the system from serving the model training needs of a large number of terminals,resulting in insufficient scalability.Secondly,the sharing of learning intermediate results during training introduces security threats,with attackers able to recover private information such as original data,leading to inadequate system security performance.Therefore,this thesis investigates the resource-constrained IoT federated partitioning learning mechanism that fuses federated learning and partitioning learning to address the above two challenges of insufficient scalability due to resource bottlenecks and insufficient security performance due to gradient sharing.The main contribution of this paper is the proposal of a scalable federated partition learning mechanism to address the challenge of insufficient scalability due to resource bottlenecks.The mechanism is optimized using the chain derivation rule of the loss function to achieve scalability under resource constraints.The split learning approach offloads computational tasks to different devices,breaking through the limitation of traditional distributed machine learning on the resources of a single computational node.The process of server backpropagation aggregates the loss functions and completes the global gradient calculation and backpropagation by chaining the derivatives of the aggregated loss functions,solving the problem of heavy server computation tasks in traditional federal partition learning.Experimental results show that the proposed scheme significantly reduces the resource overhead on the server side without affecting the model performance.To address the data security privacy risks in distributed machine learning scenarios,this paper proposes a double-mask encryption mechanism using secret sharing techniques in multi-party secure computing to achieve cryptographic protection of the original data for segmentation learning.The mechanism achieves encryption of the loss function of any client by using the secret sharing mechanism to communicate with other clients to generate random number seeds after the client computes the local loss function and generates the same mask data for encryption by a pseudo-random number generator.After the server receives the mask-encrypted data,it executes the aggregation reduction algorithm to reduce to get the global loss function,which is used for gradient derivation and completes backpropagation.This mechanism solves the possible dropout and dynamic access problems of IoT devices,and proves a good defense against gradient leakage attacks in experiments,while not affecting the model performance.In conclusion,this paper proposes a federated partition learning mechanism that fuses federated learning and partition learning to address resource bottleneck and security risk challenges in the resource-constrained massive IoT terminal scenario,which helps to improve the efficiency and security capability of future smart IoT data training.The proposed mechanism achieves scalable aggregated federated partition learning and secure aggregation of federated partition learning with double masks.