摘要：

互联网中存在大量隐私数据，因此防止网络入侵成为保护网络安全的关键问题。为提高网络入侵检测的准确率并解决其收敛慢问题，设计一种改进的堆叠自动编码器和残差网络（ISAE-ResNet）入侵检测模型。融合栈式自编码器和残差网络，首先将预处理后的数据输入到改进的栈式自编码器中，该栈式自编码器由2个副编码器和1个主编码器组成，数据经过副编码器和主编码器训练后重构出新的特征来防止过拟合问题；然后将解码层的权重捆绑到编码层进行优化，使模型参数减半来进行降维，提高模型的收敛速度；最后将处理过的数据输入到改进的残差网络中，并基于改进的ResNet网络设计一种加入软阈值函数的残差模块，通过降低数据中的噪声来提高模型准确率。在CIC-IDS-2017数据集上的实验结果表明，该模型准确率为98.67%，真正例率为95.93%，误报率为0.37%，损失函数值快速收敛至0.042，在准确率、真正例率、误报率和收敛速度方面均超过对比入侵检测模型，具有较高的有效性和可行性。

关键词: 网络入侵检测, 深度学习, 栈式自编码器, 残差网络, CIC-IDS-2017数据集

Abstract:

The vast quantities of private data on the Internet necessitate robust network intrusion prevention to safeguard network security. This study introduces an Improve Stacked AutoEncoder-ResNet(ISAE-ResNet), combining an improved self-encoder and residual network (ISAE-ResNet), to augment network intrusion detection accuracy and address the challenge of slow convergence. The model integrates an improved stack self-encoder with the Residual Network(ResNet). Initially, preprocessed data are fed into the enhanced stack self-encoder, consisting of two sub-encoders and a main encoder. By training these components, data is reconstructed with novel features to mitigate fitting issues. The advanced stack self-encoder synchronizes the weights of the decoding and encoding layers, thereby reducing model parameters by half, diminishing dimensionality, and accelerating convergence. Subsequently, the processed data is introduced into the refined ResNet, incorporating a residual module equipped with a soft threshold function to enhance accuracy by diminishing data noise. The model's efficacy and feasibility are evaluated using the CIC Intrusion Detection Systems-2017 dataset(CIC-IDS-2017). Results demonstrated a 98.67% accuracy rate, a 95.93% true case rate, a mere 0.37% false alarm rate, and rapid convergence of the loss function value to 0.042. These metrics surpass existing models in terms of accuracy, true case rate, false alarm rate, and convergence speed, thereby affirming the high validity and feasibility of the proposed intrusion detection model.

Key words: network intrusion detection, deep learning, stack self-encoder, residual network, CIC-IDS-2017 dataset