Fraud detection in financial systems is crucial for minimizing losses, ensuring compliance, and preserving user trust. However, dynamic shifts in financial networks—driven by transaction anomalies, market volatility, and data noise—pose significant challenges to detection accuracy. To address these, we propose DGNN-LNN, a novel model integrating Dynamic Graph Neural Networks with Liquid Neural Networks. The approach includes three key components: (1) a Gaussian Mixture Model to denoise raw transaction data while preserving critical features; (2) a dynamic graph construction method that captures evolving entity relationships in real-time; and (3) Liquid Neural Networks to model complex dependencies with high computational efficiency. Experiments on real-world datasets show that DGNN-LNN enhances fraud detection accuracy while reducing runtime, supporting its scalability for large financial systems.