With the increasing prevalence of graph-structured data, graph anomaly detection has emerged as a crucial research domain. Motivated by the realistic challenge that many practical problems are constrained by limited sample data, this study proposes a semi-supervised setting, unlike conventional unsupervised and supervised learning methods, where only a subset of normal samples is available. A key challenge in this context is the absence of anomalous samples, which can lead to model bias and compromise detection performance. To address this issue, we introduce a novel model, Homophily-Aware Generative Adversarial Network (HAGAN), which leverages a generative adversarial network to generate high-quality anomalous nodes. These generated nodes are seamlessly integrated into the real graph using a transformer-based graph autoencoder. Furthermore, the discriminator employs a GNN architecture enhanced with an edge homogeneity identification mechanism to improve anomaly detection. The proposed model is evaluated on four large-scale real-world benchmark datasets, and experimental results demonstrate that HAGAN consistently achieves state-of-the-art performance across multiple evaluation metrics.