<p>The effective detection of anomalies in graph data is crucial for various applications. Although existing contrastive learning methods have made some progress, they still struggle to handle diverse anomaly types concealed in the complex graph structures and the noises introduced by graph augmentations adopted. In this work, we propose a beHavior-Enhanced contrastive learning framework for Graph Anomaly Detection(HE-GAD) to address these specific challenges. We propose a novel node embedding method that guides the aggregation of node behavioral features using attribute features, maximizing the exploration of information within the graph structure. Furthermore, a new and effective contrastive learning framework is introduced to spare additional effort for devising graph augmentation methods to reduce the possible noises. Extensive evaluations with three real-world datasets demonstrate the effectiveness of our method.</p>