Detecting anomalies within complex networks is essential for identifying malicious activities, system failures, and security vulnerabilities. However, existing anomaly detection methods face two key challenges: (1) Generative Adversarial Networks (GANs) used for anomaly synthesis often fail to capture the temporal evolution of node behaviors, leading to a lack of diversity in generated anomalies; and (2) Graph Neural Networks (GNNs) often suffer from information loss due to their inherent smoothing operations, which obscure high-frequency details critical for detecting subtle anomalies. To address these challenges, we propose HADNet, a novel framework incorporating three specialized modules. The Temporal Feature Evolution Extraction Module (TFEE) captures subtle temporal changes in node features, enabling improved differentiation between normal and anomalous behaviors. The Hidden Markov Anomaly Synthesis Module (HMAS) leverages hidden Markov models to generate diverse and temporally consistent anomaly samples, enhancing the robustness of training data beyond traditional GAN-based methods. The Multiscale Time-Frequency Fusion Prediction Module (MTFFP) integrates time-domain and frequency-domain information using discrete wavelet transforms, mitigating the loss of high-frequency details caused by GNN feature aggregation. Extensive experiments on Wikipedia, Reddit, and MOOC datasets demonstrate HADNet’s superiority over state-of-the-art methods, achieving ROC-AUC improvements of 2.92%, 10.66%, and 7.72%, respectively.