Research And Implementation Of Robust Refinement Methods For Knowledge Graph

In the field of knowledge representation,data sparsity is a common challenge to be solved for large-scale knowledge graphs.In this regard,researchers have established a continuous vector space into which the triples in knowledge graph are vectorized and embedded,thus enabling distributed representation of entites and relationships.The existing models use the method of generating negative samples during training process mainly by random sampling.The quality of the negative samples generated by this method are poor,which is not significant for the robustness enhancement of the knowledge representation.This thesis combines existing knowledge representation models into generative adversarial networks to improve model performance based on existing principles and techniques related to knowledge graph-oriented representation learning.Meanwhile,with the rise of machine learning attacks,there is a risk of leakage of model parameters in the production environment,which poses a threat to system security.In this thesis,a parameter protection mechanism is designed and established to protect the model security and optimize the model efficiency through optimization algorithms.The specific research content contains the following aspects:(1)To solve the problem of low quality of negative samples generated by random sampling in the Trans series model,this thesis introduces the KB-GAT model with better singleton performance.Also inspired by generative adversarial networks,the KBGAN model is introduced to generate higher quality negative samples to improve the robustness of the model.The existing TransE and TransD models are combined with KB-GAT and tested as different combinations of generators and discriminators.The results show that the accuracy of the proposed knowledge representation framework combining KB-GAN with KB-GAT outperforms the existing mainstream single algorithms.(2)Through learning and understanding the work related to model parameter tampering attacks,this thesis proposes a parameter protection mechanism for the knowledge graph representation learning model KB-GAT from the perspective of defense against model tampering attacks,and designs an efficient protection framework based on parameter criticality.According to the criticality level,the model parameters are verified in a hierarchical manner by cryptographic verification before system deployment and during operation to ensure that the model parameters are not tampered with and thus enhance the robustness of the knowledge graph.Finally,the experimental results are analyzed to demonstrate the effectiveness of the protection mechanism on the stable operation of the model.(3)The model parameter protection mechanism added in the actual deployment increases the resource and time overhead of the system,which needs to be constrained and optimized.This thesis transforms the protection mechanism into a constraint optimization problem and establishes a security model,while adding execution constraints and duration constraints to the KB-GAT protection mechanism.This thesis proposes the use of the fruit fly algorithm(FOA)to obtain the near-optimal solution,and also uses the MFOA algorithm,which is a multigroup policy improvement of the fruit fly algorithm,in order to address the limitation that the fruit fly algorithm cannot ensure the global optimal solution.The experimental evaluation shows that the multigroup FOA algorithm(MFOA)is effective for improving the efficiency of the protection mechanism.