Cross-Domain Few-Shot Learning (CD-FSL) has gained considerable attention in recent years due to its potential to transfer knowledge from a source dataset to a target dataset with minimal data. However, the significant domain discrepancies between source and target datasets present substantial challenges. Although Graph Neural Networks (GNNs) have shown impressive progress in few-shot learning tasks, they often suffer from overfitting problems, which hinders their scalability and cross-domain generalization. To address these challenges, we propose a GNN framework for CD-FSL that incorporates Equiangular Embedding (EE) and Dynamic Adversarial Augmentation (DAA). This framework leverages learnable equiangular label embeddings to capture relationships within the data and employs dynamic adversarial augmentation to enhance cross-domain generalization. We trained our model on the mini-ImageNet dataset and conducted comprehensive experiments on four target datasets: CropDiseases, EuroSAT, ISIC, and ChestX. Our approach can be integrated with existing CD-FSL methods to improve performance. Remarkably, our 5-way 1-shot fine-tuning experiments using the EE method achieved an average accuracy of 50.78% (+6.08%) across four target datasets. This performance is comparable to that of a meticulously designed adversarial enhancement method. Our experimental results highlight the substantial improvement our method brings to CD-FSL performance, achieving a new state-of-the-art level.