<p>Existing Intrusion Detection Systems (IDS) for wireless networks typically rely on traditional or expert-driven feature selection methods. It is essential to choose only the relevant or correlated features through feature selection, which can reduce learning time and improve the performance of IDS. Feature optimization is needed to handle imbalanced datasets. Existing IDS approaches often struggle with high-dimensional data and redundant features, which can reduce predictive performance and increase computational cost. To address these limitations, we propose a Rough Set-based feature selection framework that achieves high accuracy while minimizing the number of features and ensuring well-calibrated predictions in IDS for 802.11 wireless networks, aiming to improve feature utilization. Rough Sets efficiently reduce dimensionality and identify the most relevant features, improving the effectiveness of the classifier. Hybrid and ensemble approaches are needed in IDS to improve detection and reduce the gap between system understanding and real-world attacks. The proposed work studies the performance of a hybrid classifier consisting of autoencoder(AE) stacked upon multilayer perceptron (MLP) over the proposed roughset-driven features specifically for AWID3 dataset. To understand the performance of the proposed hybrid AE-MLP classifier, we prepared experimental setup with two variations of it: (i)proposed single-layer AE-MLP (SAE-I) and (ii) proposed stacked AE-MLP (SAE-II). To further evaluate the generalizability of the proposed approach, we applied it to the AWID2 dataset to test its performance.The roughset-based feature selection method selected common 108 features from AWID3 dataset and 89 features from AWID2 dataset. Based on rank dependency, we selected top 13 features with positive rank dependency from AWID3 and top 5 features from AWID2 dataset. These features are optimum as per the proposed work since these are the possible near-smallest feature sets prepared for our proposed pipeline. The subsets of these features (viz., 12,9,7 for AWID3)are falling under the reduced feature sets. The main objective to identify the essential optimum number of features sets for the proposed roughset driven autoencoder based hybrid classifiers utilizing mainly the AWID3 and further checking its generalizability future scope on other datasets beyond AWID3. The proposed autoencoder based classifier SAE-I achieves 99% accuracy, precision, recall, and F1-score, with an AUC of 0.999625 and an overall detection accuracy of 98.625% using 108 common features in the AWID3 dataset. SAE-I exhibits an accuracy of 99%, with an AUC of 0.9992 and an overall detection accuracy of 98.714% with 13 optimal features, demonstrating its effectiveness in the AWID3 dataset. The proposed SAE-II is the better performer in terms of AUC(99.975%) and overall detection accuracy of 98.828% than SAE-I using 13 features. The Rough Set-based subset was compared within the proposed SAE-I pipeline alongside PCA, Mutual Information, XGBoost, PSO, and XGBoost+PSO, achieving 99% accuracy with only 13 features and providing a better balance of predictive performance, feature reduction, and well-calibrated predictions than the other state-of-the-art methods using the AWID3 dataset. In addition to this, we compared the performance of our proposed classifier with baseline models such as Random Forest and MLP to analyze its performance relative to baseline models with AWID3. The optimal 5-feature set obtained from the AWID2 dataset shows an AUC of 95.69%, with a detection accuracy of 81.68% for SAE-I, compared to SAE-II, which shows a marginal difference.</p>