In this paper, we propose a deep learning-based framework for cyberattack detection and classification for Windows-based Internet of Things (IoT) in a smart city. The smart city leverages the IoT to improve services including health, safety, and traffic while also increasing the efficiency of urban living. In smart cities, one of the most widely used IoT technologies is Windows-based systems. Moreover, Windows IoT platforms provide a flexible environment for interacting with and operating IoT devices. However, due to weak authentication and insufficient access control in IoT, the infrastructure of smart cities becomes vulnerable to various cyberattacks. Deep learning techniques can efficiently enhance cybersecurity in smart cities. There are very few works on deep learning techniques for detecting cyberattacks in Windows-based IoT systems within smart cities. In this research, we developed powerful Autoencoder-based bidirectional long short-term memory (BiLSTM) and stacked Autoencoder-based dense neural network (DNN) hybrid models to investigate our proposed framework for detecting cyberattacks in the fog layer of smart cities. The autoencoder-based model is highly effective and performs dimensionality reduction and feature extraction. These models can detect instances of cyberattacks using binary classification and classify the type of cyberattack using multiclass classification. After conducting our experiment on Windows-based IoT data, we identified 7 cyberattacks using the models. In multiclass classification, the stacked Autoencoder-DNN achieved the highest accuracy of 99.86%. In binary classification, the stacked Autoencoder-DNN also outperformed with the highest accuracy of 99.90%. The stacked Autoencoder-DNN also achieved an outstanding accuracy of 99.99% in detecting man-in-the-middle attacks.