<p>Fraud detection in graphs has attracted considerable attention in various domains, where a primary challenge is the data heterophily. Despite the success of existing GNN-based methods on handling heterophily on static graphs, their performance on dynamic graphs are suboptimal due to not fully utilizing two crucial types of temporal information in dynamic graphs: the temporal paths and time spans. To address this problem, we present <Emphasis Type="Underline">T</Emphasis>emporal <Emphasis Type="Underline">P</Emphasis>ath <Emphasis Type="Underline">A</Emphasis>ggregator (TPA), a novel dynamic graph fraud detector, which is capable of modeling heterophilic data while capturing the temporal information by sampling and aggregating temporal paths associated with each node. A time span encoder module is also devised to capture the fine-grained time span information. Results of extensive experiments demonstrate that our model outperforms state-of-the-art baselines.</p>