Nowadays, smart Internet of Things (IoT) technology as a new paradigm has been widely used in different fields of our lives. Massive amounts of high-dimensional time-series data are generated by sensors in the smart IoT, where data anomalies may often occur. Detecting anomalies in sensor data is crucial to ensure IoT systems’ stability, security, and reliability. However, the traditional anomaly detection methods often entail transmitting the data to the cloud or a centralized server for processing, which can result in increased latency, making it unsuitable for real-time application scenarios. Conversely, deploying anomaly detection directly on edge devices improves detection speed and efficiency. To this end, this paper proposes a lightweight time-series anomaly detection model combining Attention Anomaly Detection Generative Adversarial Networks (att-ADGANs) and Knowledge Distillation (KD). In the initial phase, we train a GAN network with encoder and decoder architectures to capture normal data distribution features. Subsequently, the GAN network is compressed and distilled into a lightweight model using KD. Furthermore, we present a Series-Knowledge Distillation GAN (S-KDGAN) framework that leverages progressive distillation to enhance distillation efficiency and minimize performance degradation. Experimental results show that the S-KDGAN compresses the computation at ratios of 2.37:1, 48.85:1, and 15.93:1 on the SWMRU, KDDCUP99, and HomeC datasets, and performance degradation rates of only 1.38%, 2.27%, and 1.61%. After comparing with other models, the effectiveness and superiority of the S-KDGAN distillation framework are proved.