<p>Intrusion detection in computer networks is critically important, as it directly impacts connectivity reliability and overall security. However, detecting network intrusions remains a challenging task due to the rapid evolution of cyber threats, increasingly sophisticated attack techniques, and the requirement for real-time monitoring. In addition, many state-of-the-art learning models demand large volumes of labeled data to achieve effective threat detection. Although numerous network intrusion detection approaches have been proposed in recent years, they continue to face significant challenges arising from the continuous emergence of novel threats that are not effectively detected by existing systems. To address these issues, this work proposes an end-to-end deep learning-based intrusion detection framework. In the data preprocessing stage, categorical features are transformed into binary vectors using one-hot encoding, followed by normalization to ensure uniform feature scaling. Subsequently, a convolutional layer is employed to capture the local spatial patterns present in network traffic data. Temporal dependencies and sequential characteristics of traffic flows are then modeled using a Bidirectional Long Short-Term Memory (Bi-LSTM) network, where network packets are sequentially fed into the Bi-LSTM to generate representative packet-level feature vectors. Furthermore, the Adaptive Cheetah Optimization (ACO) algorithm is utilized to optimize the hyperparameters of the Bi-LSTM model, thereby enhancing learning efficiency, convergence speed, and detection performance. Finally, a SoftMax classifier is used to categorize network traffic into normal and intrusion classes. Experimental results demonstrate the effectiveness of the proposed approach, achieving an accuracy of 98.50%, precision of 98.48%, recall of 98.59%, and an F1-score of 98.54%.</p>