Research on Threat Attribution Mechanisms and Early Warning Strategies for Federated Models in Secure Educational Data Sharing

Abstract

As educational digital transformation accelerates, the demand for cross-institutional educational data sharing has become increasingly urgent. Federated learning (FL), characterized by its privacy-preserving paradigm of keeping data local while sharing model updates, has emerged as a critical framework for secure educational data sharing. However, FL in educational settings faces multifaceted security threats including gradient inversion, model poisoning, and backdoor attacks, while existing defense mechanisms exhibit significant shortcomings in threat traceability and early-warning response—particularly in dynamic modeling of temporal behavioral evolution of participating nodes. This paper proposes TKG-FedAlert, a threat traceability mechanism and dynamic early-warning strategy framework integrating temporal knowledge graphs (TKG) into federated educational model security. The framework constructs a temporal knowledge graph of educational federated node behavioral patterns, enabling structured modeling of gradient update dynamics, behavioral sequences, and inter-node relationships. A temporal graph attention network (TGAT) combined with temporal anomaly detection achieves precise malicious node traceability and threat warning, while a tiered response mechanism supports real-time intervention. Simulation experiments on realistic multi-university FL scenarios demonstrate that TKG-FedAlert significantly outperforms existing baseline methods in threat traceability accuracy (84.2%), early-warning F1 score (88.4%), and system response latency (1.9 rounds), providing both theoretical foundations and practical pathways for building trusted federated educational data sharing ecosystems.