Graph Neural Network (GNN) models are increasingly being applied in fraud detection, which involves identifying fraudsters in graphs. GNN-based methods on static graphs ignore the important temporal information, resulting in low accuracy in time-sensitive applications. GNN-based methods on dynamic graphs take into account the temporal information but overlook the influence differences between long-term and short-term temporal information, limiting detection accuracy. To address these issues, we propose STM: a Spatio-Temporal Model for fraud detection on dynamic graphs. We split original graphs into a set of snapshot graphs based on time frames. From intra-perspective, in each snapshot graph, we present a differential aggregation method based on spatio-temporal distance to effectively aggregate both the spatio and short-term temporal information. From the inter-perspective, we capture the long-short term temporal information among all snapshot graphs based on self-attention mechanism. Extensive experiments on real-world datasets demonstrate that our method effectively and efficiently detects fraudsters in dynamic graphs. Our code can be found at https://anonymous.4open.science/r/STM/ .