As network attack techniques become increasingly sophisticated and stealthy, traditional detection methods are struggling to cope with emerging threats. This paper proposes a complex network attack detection strategy based on a multi-layer Graph Neural Network (GNN), which incorporates adaptive graph structure modeling and multi-scale feature aggregation techniques. By dynamically adjusting the network topology and capturing multi-level attack features, the proposed strategy enables precise identification and classification of complex attack behaviors. Extensive experiments on multiple public datasets demonstrate that this approach significantly outperforms traditional detection techniques, particularly in handling advanced persistent threats (APT) and other sophisticated attack scenarios, exhibiting strong adaptability and robustness. The findings suggest that this detection strategy not only expands the theoretical application of GNNs in network security but also provides an innovative technical solution to addressing complex network attacks in practice.