Computer Network Security System Optimization Based on Improved Neural Network Algorithm and Data Search

Frequent hacker attacks and network viruses pose a serious threat to network security, and traditional static and passive defense technologies can no longer meet the high security demands of networks. This paper proposes a computer network security intrusion detection algorithm based on the optimization of convolutional neural networks through transfer learning. Initially, a one-dimensional convolutional neural network (1D-CNN(T)) is trained to form a stable model. Subsequently, in the transfer learning phase, a target function is constructed, experimental data is acquired, and 1D-CNN(E) is trained. The network parameters are fine-tuned using a co-gradient algorithm, and features are extracted from the experimental data, with connections being dropped during the process to reduce overfitting. Based on this algorithm, an intrusion detection system is designed, which is modularly constructed to provide a complete intrusion detection solution. The paper also provides a detailed analysis of modules such as data acquisition, preprocessing, feature extraction, and neural network classifiers. Finally, the performance of the proposed convolutional neural network-based intrusion detection algorithm, which leverages transfer learning, is validated through a neural network model experiment. This algorithm is compared with other classical classification algorithms commonly used in intrusion detection, including RF, AlexNet, LeNet-5, CNN, and BiLSTM, each serving as a classifier. The experimental results demonstrate that the proposed classification algorithm achieves an accuracy of 83.58% and a recall rate of 84.49%. Compared to the RF algorithm, the proposed method exhibits improvements of 8.87% in accuracy and 9% in recall rate. When benchmarked against the AlexNet model, the accuracy and F1-Measure of the proposed algorithm are enhanced by 6.56% and 7.26%, respectively. The classifier in TLCNN-IKNN achieves the highest classification accuracy using weighted Euclidean distance, with an accuracy rate of 99.07% for detecting COMBO attacks in the Bashlite botnet, 97.02% for detecting Junk attacks, 97.71% for detecting Scan attacks, 98.08% for detecting TCP attacks, and 100.00% for detecting UDP attacks. These findings underscore the effectiveness of the transfer learning-based intrusion detection algorithm in boosting the detection rate of intrusion detection systems and reducing false positive rates, thereby establishing its high practical application value.