基于CA-ResNet网络与nadam优化的入侵检测算法

Intrusion Detection Algorithm Based on CA-ResNet Network and Nadam Optimizer

【目的】针对深度学习模型在网络入侵检测中进行参数训练时因梯度消失而导致深度学习模型过拟合在测试集上准确率下降的问题。提出一种结合LeakyRelu激活函数与ResNet的网络入侵检测算法,即CA-ResNet,并采用nadam优化器对模型进行优化。【方法】该模型在DNN的基础上增加了网络的层次,结合了ResNet和LeakyRelu激活函数。【结果】解决了模型训练时梯度消失的问题,保证了该模型在测试数据集上的表现,使得训练的模型的泛化能力更强,同时通过增加网络的单层维度和总层次的深度,提高了网络的特征提取能力和对尺度的适应性。【结论】使用KDD Cup99数据中的Corrected数据集对算法进行验证。实验表明,该算法与CNN和CNN-BiLSTM算法相比具有更高的准确率和F1-score,准确率能够达到95.0%,F1-score能够达到97.5,时间复杂度为线性时间复杂度。

[Purposes]For deep learning model in network intrusion detection parameter training is easy to cause the gradient disappeared and further cause deep learning model fitting of the decline of the accuracy on the test set.A CA-ResNet intrusion detection algorithm based on Residual Network(ResNet)incorporating LeakyRelu activation function was proposed,and the model is optimized by nadam optimizer.[Methods]This model adds the layer of network to DNN,through the combination of residual network and LeakyRelu activation function.[Findings]The problem of gradient disappearance during model training is solved.The performance of the model on the test data set is ensured,and the generalization ability of the trained model is enhanced.At the same time,the feature extraction ability and adaptability to scale of the network are improved by increasing the single-layer dimension and the depth of the total layer of the network.[Conclusions]The algorithm is verified by the corrected dataset in KDD Cup99 data.Experimental results show that the algorithm has higher accuracy and F1-score than CNN and CNN-BiLSTM algorithm,the accuracy rate can reach 95.0%,F1-score can reach 97.5%,the time complexity is linear time complexity.