Noise-adaptive correction for robust graph neural networks in trusted graph computing

Trusted computing systems and blockchain-enabled security applications increasingly rely on Graph Neural Networks (GNNs) for trust graph analysis, fraud detection, and anomaly identification. In these security-critical deployments, graph data is routinely subject to adversarial manipulation—including Sybil attacks, attribute poisoning, and label flipping—making robustness a fundamental system-level trust requirement. While GNNs achieve strong performance on homophilic graph data such as citation networks, in compound noise environments where structural and feature noise are combined, attention mechanisms become distorted and performance degrades severely. Existing studies either rely on structure learning that requires high computational cost of O (N²) or need clean validation data, limiting practical applicability in real-world trusted computing deployments. In this paper, we propose NAC (Noise-Adaptive Corrector), a framework that leverages homophily properties to simultaneously achieve computational efficiency and robustness. Inspired by trusted computing principles, NAC employs a dual-path architecture that separates a trusted reference signal path from an observed noisy path, using KL divergence to quantify trust deviation at the node level. NAC actively detects and corrects noise without label information by minimizing the KL divergence between reference signals generated through neighbor averaging and observed signals. Furthermore, we introduce a buffering strategy that omits the Reference Encoder computation during inference, reducing actual computational load by approximately 50%. Experimental results on various benchmark datasets show that the proposed NAC-Practical achieves 6. 2%p improved accuracy over baseline models without requiring clean original data, demonstrating superior robustness and establishing a foundation for trustworthy GNN deployment in blockchain-enabled security systems.