<p>As the Internet of Things (IoT) continues to expand, the proliferation of smart devices has resulted in dynamic network structures, creating new challenges for node anomaly detection. This paper proposes Anomal-EFD, a self-supervised dynamic IoT anomaly detection model. First, it employs a random sliding window mechanism to dynamically adjust the time window size, better adapting to varying network rhythms over time. The gradient aggregation mechanism captures both short-term and long-term changes. Second, a multi-head attention mechanism enables precise aggregation of neighboring node information, significantly enhancing feature fusion. Finally, an unsupervised module combines self-supervised learning with contrastive learning to achieve node anomaly detection. Our source code available at: <a href="https://github.com/decentbb/Anomal-EFD">https://github.com/decentbb/Anomal-EFD</a>.</p>