The rapid growth of Internet of Things (IoT) devices has expanded the cyber attack surface, posing a challenge to IoT security. Some deep-learning-based detection methods have been designed to detect malicious attacks in the IoT by analyzing network traffic. However, the algorithm computational complexity of existing methods is usually high due to having a large number of parameters and iterative training inference, making them difficult to implement on IoT gateways which have limited computational and storage resources. To this end, this article proposes a lightweight malicious traffic detection model for IoT based on lightweight residual block (LRB) modules, named LightGuard. Specifically, the LRB module designs a unique residual structure based on the construction idea of ShuffleNetV2, which enables LightGuard to achieve high detection performance while reducing the parameters, computations, and inference time of the model. In addition, the LRB module replaces the traditional convolution with a lightweight convolutional module called the ghost module to generate feature maps at low cost while without compromising detection performance. We evaluate the effectiveness of LightGuard by comparing it with seven advanced baseline models on four generic data sets. The experimental results show that LightGuard achieves more than 99.6% accuracy on all four data sets, while exhibits significant advantages with low computational complexity.