Fingerprint recognition is a widely used biometric technology due to its uniqueness and reliability. Traditional fingerprint representation methods, often based on manually designed rules, struggle with robustness under varying fingerprint quality and noise. Recent approaches leveraging convolutional neural network (CNN) and transformer have improved performance but fail to capture the inherent topological structures of fingerprint minutiae effectively. To address these limitations, we propose a novel fingerprint representation learning method based on Graph Neural Network (GNN). Fingerprint images are transformed into graph structures using minutiae as nodes and edges constructed via the K-nearest neighbors algorithm. To mitigate noise and structural redundancy, we introduce an attention-based subgraph extraction mechanism guided by the information bottleneck principle, enabling the generation of robust and discriminative fingerprint embeddings. Experiments on multiple benchmark datasets demonstrate the superior performance of our method compared to existing approaches, achieving higher accuracy and robustness. Ablation studies validate the effectiveness of the proposed components, providing a promising solution for efficient and reliable fingerprint recognition.