An intrusion detection system serves as a critical barrier in network security protection. Existing intrusion detection approaches face several challenges when processing high-dimensional and sparse network traffic data, including difficulties in feature extraction, data category imbalance, and a lack of self-optimization capabilities. To address these issues, this study proposes an intrusion detection method based on an improved differential evolution algorithm and stacked autoencoder, termed iDE-SAE. A novel Geometric SMOTE algorithm is employed to adaptively interpolate minority class samples, thus enhancing their feature representation and mitigating the adverse effects of class imbalance. A stacked autoencoder is then utilized to extract essential low-dimensional features from the high-dimensional data. Based on these extracted features, a deep neural network classifier is trained, and the classifier’s hyperparameters are automatically optimized using the proposed improved differential evolution algorithm to achieve optimal classification performance. Experiments conducted on the CIC-IDS2017 and UNSW-NB15 datasets demonstrate classification accuracies of 99.66% and 99.61%, respectively, with F1-scores of 99.40% and 99.65%. These results validate the effectiveness of the proposed method in intrusion detection tasks. High detection accuracy is achieved while also enabling autonomous hyperparameter tuning.