Anomaly Intrusion Detection Systems (AIDS) in the Internet of Things (IoT) environments have been revolutionized by the beginning of Artificial Intelligence (AI) techniques. For example, deep learning (DL) models and transformer models can sift through and learn from vast amounts of big data to detect complex and complicated intrusion patterns. The current work presents a TabNet model tuned with Google OSS Vizier-hyperparameters tuning service-to increase the model’s performance for AIDS tasks. The TabNet showed great promise due to its detection and capacity to handle numerical and categorical data. We implemented experimentation based on benchmark datasets, NSL-KDD, CICIoT2023, and RT_IoT2022 and the model has evaluated its efficacy in both binary and multi-classification types. The experimental results of our evaluation in binary classification are 99.9%, 99.92%, and 99.16% on NSL-KDD, CICIoT2023, and RT_IoT2022 datasets, respectively. While for multi-classification, they are 98.29%, 98.43%, 98.10% for NSL-KDD, CICIoT2023, and RT_IoT2022 datasets, respectively. Moreover, this work also incorporates using of Explainable AI (XAI) techniques, specifically the SHapley Additive exPlanations (SHAP) function to illustrate the contribution of individual features to the model’s predictions and decisions. The combination of perceptive explanations of XAI with the optimized TabNet’s powerful feature learning presents a major advancement in the development of highly effective, and interpretable AIDS.