基于辅助熵减的神经常微分方程入侵检测模型

Auxiliary Entropy Reduction Based Intrusion Detection Model for Ordinary Differential Equations

为了提高深度学习模型入侵检测任务的检测效率和分类的准确性,文章提出一种基于辅助熵减的神经常微分方程(E-ODENet)入侵检测模型。该入侵检测模型通过参数常微分方程定义连续的隐藏状态,不需要再分层传播梯度与更新参数,减少了内存的消耗,极大地提高了效率。使用信息瓶颈进行特征降维,提取与分类任务相关的主要信息,同时使用标签平滑和熵减损失来提高模型的泛化能力和准确性。在NSL-KDD数据集上进行训练和测试,测试得到的检测准确率为99.76%,证明该模型优于其他入侵检测模型。

In order to improve the detection efficiency and classification accuracy of the intrusion detection task of deep learning model,this paper proposes an intrusion detection model based on the auxiliary entropy subtraction of the divine ordinary differential equation(E-ODEnet).This intrusion detection model defines continuous hidden states by parametric ordinary differential equations,while does not require further hierarchical propagation of gradients with updated parameters,while reducing memory consumption and greatly improving efficiency.Feature dimensionality reduction is performed using information bottlenecks to extract the main information relevant to the classification task,while label smoothing and entropy reduction loss are used to improve the generalization ability and accuracy of the model.This experiment is trained and tested on the NSL-KDD dataset,and the accuracy rate of the experimental result is 99.76%,which is better than other intrusion detection models.