CyberVeriGNN: A Graph Neural Network‐Based Approach for Detecting Fake Cyber Threat Intelligence

ABSTRACT Cyber threat intelligence (CTI) is vital for the proactive identification and mitigation of cybersecurity threats. However, inconsistent quality and the spread of misinformation undermine decision‐making, weaken defensive measures, and increase exposure to attack. To address these challenges, we propose CyberVeriGNN, a graph neural network (GNN) framework that detects forged CTI by jointly modeling semantic consistency and structural coherence. The method parses unstructured threat reports into heterogeneous attack subgraphs, enriches node and edge representations using BERT contextual embeddings and MITRE ATT&CK semantics, and fuses these features via a multi‐head graph attention network (GAT). A graph autoencoder (GAE) subsequently computes reconstruction error, yielding an interpretable anomaly score. Evaluated on real CTI corpora augmented with expert‐validated adversarial samples, CyberVeriGNN achieves a precision of 81.57% and an F1 score of 81.25%, significantly outperforming both traditional and graph‐based baselines. These results demonstrate the value of deep semantic–structural fusion for CTI authenticity verification and lay a strong foundation for scalable, trustworthy validation in security‐critical environments.