Carbon Fiber Reinforced Plastic (CFRP) is extensively used in space transportation systems owing to its high strength-to-weight ratio. However, internal defects such as interlaminar delamination can critically degrade structural integrity, necessitating reliable defect localization for reusable space vehicles. Infrared stress measurement has emerged as a promising non-destructive testing (NDT) technique, as it provides full-field distributions of the sum of principal stresses on the surface (DSPSS), from which subsurface defects may be inferred. In this study, a previously proposed defect localization framework based on graph neural networks (GNNs) is applied to CFRP interstage structures with holes, with emphasis on robustness under experimentally motivated noise conditions. Synthetic datasets were generated using finite element analysis (FEA) under tensile loading, in which interlaminar delamination defects were explicitly modeled by Teflon inserts between selected plies. To mitigate the dominant influence of stress concentrations around openings, difference-normalized DSPSS fields were employed as input features. The FEA surface mesh was represented as a graph, where each surface element was treated as a node with spatial coordinates and normalized DSPSS values as node features. A Graph Attention Network (GAT) was adopted to adaptively weight neighboring nodes, enabling effective learning from highly non-uniform stress fields caused by defects and geometric discontinuities. The defect localization task was formulated as a multi-class classification problem to estimate both in-plane defect regions and corresponding insertion layers. The results indicate that the proposed framework can localize defects in perforated CFRP structures while maintaining robustness against noise and suppressing false detections in defect-free regions. This study establishes a scalable foundation for future application to experimental infrared stress measurement data and practical structural health monitoring of aerospace composite structures.