FEDERATIVE GNN-XAI MODEL FORPREDICTING COMPROMISE OF ACCOUNT RECORDS IN ZERO TRUST ENVIRONMENT

Abstract

The article presents a methodological approach to developing an intelligent system for predicting user account compromise in corporate information environments. The proposed system integrates federated learning, graph neural networks, and explainable artificial intelligence within the Zero Trust concept, ensuring an enhanced level of security for authentication and access management processes through decentralized data processing and privacy preservation during collaborative model training. One of the key features is local model training without transferring primary data to a central repository, which eliminates the possibility of interception or unauthorized access. Aggregation of local updates is performed using federated optimization mechanisms that account for the heterogeneity of data from various corporate domains. The graph module formalizes inter-user and inter-system relationships as a directed graph, allowing for the identification of latent behavioral dependencies and potential compromise risks at the level of individual connections between authentication objects. The integration of an explainable artificial intelligence component ensures transparency in the decision-making process, formalizes the justification of predictions, and reduces the frequency of false positives. The system is implemented in a Zero Trust paradigm, which involves continuous verification of user and device actions regardless of the trust level, ensuring adaptive real-time anomaly response. The obtained results demonstrate an increase in the accuracy of predicting account compromise compared to traditional centralized machine learning models and a reduction in the frequency of false positives. The proposed approach can be used to build adaptive security monitoring systems in critical information infrastructures operating in highly dynamic and distributed environments.