Abstract In intrusion detection, imbalanced datasets often degrade the accuracy of certain detections, potentially allowing malicious traffic to remain undetected and leading to significant losses. The multi-layer perceptron (MLP) offers distinct advantages for intrusion detection, as attack patterns often follow complex, nonlinear relationships. These patterns can be effectively captured through MLP’s multiple nonlinear transformations, such as ReLU and Sigmoid activation functions, which are especially beneficial for intrusion detection. Additionally, MLP exhibits low resource consumption, making it suitable for resource-constrained environments. However, MLP often struggles to accurately classify minority classes in imbalanced datasets due to its limited feature extraction capabilities. In contrast, convolutional neural networks (CNNs), particularly AlexNet’s small convolutional filters, offer more precise feature extraction for detailed dataset features. Therefore, this study integrates AlexNet’s feature extraction module with MLP and incorporates the SKNet attention mechanism to improve the recognition of minority classes. Experimental results show that our enhanced MLP algorithm outperforms the standard MLP across all seven proposed classification tasks. Specifically, the F1 scores for BotnetARES and PortScan show significant improvements of 18.93% and 26.57%, respectively, validating the efficacy of the algorithmic enhancements.