Graph-based fraud detection is crucial in domains like social networks and finance, posing a challenge in identifying fraud activities. Existing methods primarily focus on the link patterns between nodes in the spatial domain, while placing greater emphasis on extracting information across different frequencies in the spectral domain. However, they overlook the fact that fraudulent and normal samples exhibit a high degree of similarity due to the presence of fraud camouflage. To alleviate this problem, this paper proposes an adaptive masked graph neural network (AdaMask-GNN). AdaMask-GNN first adaptively masks the neighbor features of the center node in the message propagation stage to achieve feature augmentation and adaptive denoising. Then, in the message aggregation stage, the center node features are effectively combined with the context structure features through decomposition and aggregation operations. Finally, the model classification feature space is optimized by introducing asymmetric feature consistency contrast loss. Extensive empirical evaluations on three benchmark datasets substantiate the efficacy of our method in advancing graph-based fraud detection, while simultaneously promoting a more discriminative and well-structured feature space for classification.