With the popularization of emerging technologies such as 5G and IoT, network attacks present the characteristics of cross-platform penetration, enhanced covertness, and intelligent attack chain. Aiming at the problems of insufficient feature characterization ability and weak dynamic adaptability faced by the traditional rule-base-based intrusion detection methods, combining with deep learning, The architecture diagram of the network intrusion detection model was established based on the concept of network intrusion detection. The four challenges faced by current typical deep learning-based models—such as CNN(convolutional neural network), RNN(recurrent neural network), GAN(generative adversarial network), and AE(autoencoder)—were summarized and analyzed. Accordingly, the future development direction was analyzed and prospected.
YANG Yu, YUAN Jin-liang, GAO Min-na, SHEN Fang, GUO Rui, WANG Ming-qi.
Systematic Review of Intrusion Detection Based on Deep Learning[J].
Science Technology and Engineering, 2026, 26(2): 450-464 DOI:10.12404/j.issn.1671-1815.2501499
NSL-KDD数据集的主要特点在于其精炼后的数据结构和丰富的网络流量特征。相较于原始的KDD Cup 1999数据集,NSL-KDD消除了冗余记录,改善了类别不平衡问题,从而提高了模型的训练效率和检测准确性。此外,该数据集包含了多种网络协议和攻击类型的详细信息,使得其在入侵检测领域的应用具有广泛的代表性和实用性。
Kyoto 2006+数据集是一个由日本京都大学创建的网络入侵检测数据集。这个数据集主要用于评估网络入侵检测系统(NIDS)的性能,数据集包含了多种类型的攻击,包括对Web服务器、数据库和其他网络服务的攻击。此外,数据集包含93 076 270个会话,提取了24个特征,包括14个统计特征和10个额外特征。包含14个统计特征,源自KDD Cup 99数据集,以及10个额外特征,用于NIDS的进一步分析和评估。
CICIDS2017数据集是一个由加拿大网络安全研究所(Canadian Institute for Cybersecurity,CIC)发布的网络流量数据集,它被广泛用于网络流量分析和入侵检测研究,包含良性流量和7种常见攻击类型的网络流量,数据集完全标记,并使用CICFlowMeter软件提取了超过80种网络流量特征。
卷积神经网络(CNN)的概念最早可追溯至1989年LeCun等[36]的开创性工作。该研究首次系统阐述了CNN的多层卷积结构与反向传播训练方法,并在手写数字识别任务中验证了其有效性。这一里程碑式成果不仅奠定了CNN在图像识别领域的理论基础,更为后续计算机视觉技术的突破提供了核心框架。传统CNN的核心优势在于其局部感知与权值共享机制,能够通过相关滤波器自动捕捉图像中的空间依赖关系[37],从而显著降低数据预处理复杂度。典型的LeNet、AlexNet等架构已在物体检测、语义分割等任务中展现出优越性能。近年来,研究者逐步将CNN的强特征提取能力拓展至网络安全领域,特别是在网络入侵检测场景中实现了创新应用。文献[38]提出改进型卷积神经网络(ICNN)用于无线网络入侵检测,旨在解决传统入侵检测技术在无线网络环境中存在的高误报率、低检测效率和泛化能力不足的问题。该方法通过跨层聚合网络设计和Dropout正则化等技术,增强了模型的特征提取能力和泛化能力,并利用深度学习的端到端学习特性,自动提取样本特征,简化了入侵检测的实现过程,包含多个卷积层、池化层和全连接层,能够有效提取无线网络流量数据中的关键特征。通过跨层聚合和Dropout技术,模型在训练过程中能够避免过拟合,提高泛化能力。此外,ICNN采用随机梯度下降算法进行优化,能够逐步调整网络参数,直至模型收敛,在KDDTest+数据集上的检测准确率较传统LeNet-5和DBN模型分别提升8.82%和0.51%,同时误报率显著降低,这一突破验证了CNN处理非图像结构化网络数据的潜力,但该模型在训练过程中可能会出现梯度弥散和局部最优解的问题,影响了模型的训练效果和泛化能力,其次,对于样本数量较少的攻击类型(如U2R),ICNN的检测准确率相对较低。这表明在处理不平衡数据集时,模型的性能会受到一定影响。为进一步解决特征选择偏差与零日攻击识别难题,文献[39]设计了CNN与深度分水岭自动编码器(deep watershed autoencoder,DWA)的混合架构(CNN-DWA)。在特征选择、数据不平衡处理和零日攻击检测方面具有显著优势。CNN-DWA模型通过卷积层和池化层自动提取特征,并利用动态权重分配机制优化数据不平衡问题,显著提升了模型的检测性能。实验结果表明,CNN-DWA在KDD CUP 1999数据集上的准确率达到99.9%,显著优于传统CNN模型(94.54%),但CNN-DWA需要较高的计算资源,尤其是在处理大规模数据集时,并且由于结构较为复杂,训练和测试过程所需要的时间远高于一般模型。在模型优化层面,文献[40]引入SiLU激活函数与批量归一化技术,通过归一化每个小批量(mini-batch)的数据,减少内部协变量偏移(internal covariate shift),从而加快模型的收敛速度,SGD通过随机选择小批量数据进行梯度计算,逐步更新模型参数,以最小化损失函数,训练深度学习模型时具有良好的收敛性能和泛化能力,使CNN在KDD Cup 99数据集上经过10轮训练即达到99.26%的平均准确率。文献[41]通过融合混合自编码器与残差时序网络的协同架构,实现了高维网络流量数据的高效特征提取与精准分类。在特征选择阶段,采用CNN-GRU混合自编码器进行深度降维:首先利用卷积层结合双池化策略提取流量的空间局部模式(如异常协议字段组合),通过非线性编码函数融合池化特征;随后将空间特征序列化输入GRU单元,利用其重置门与更新门建模攻击行为的时序依赖性(如DDoS流量脉冲),最终通过重建损失函数优化特征压缩过程,显著消除冗余。在分类阶段,设计混合残差LSTM(HRL)架构:通过三层残差模块(含跳跃连接)解决深层CNN的梯度消失问题,增强对复杂攻击模式(如多态恶意软件)的空间表征能力;引入多头自注意力机制动态加权关键特征区域(如攻击载荷偏移地址);并采用LSTM门控单元精准建模长周期攻击行为,最终将ResNet的空间特征图与LSTM的时序状态拼接,经全连接层融合后由Softmax分类器输出。该模型在NSL-KDD等数据集上达到95.7%以上准确率,尤其擅长处理高维时序流量,但其计算复杂度较高,且对样本稀疏的少数攻击敏感,适用于云端部署场景,未来需结合联邦学习优化零日攻击检测能力。
在当今数字化时代,网络入侵检测系统(IDS)已成为保障网络安全的关键防线。随着中国网络技术的迅猛发展,各行业对网络安全的需求日益增长,尤其广泛应用于物联网(internet of things, IoT)、车联网及软件定义网络(software defined network,SDN)等新兴网络架构中。由于上述网络在资源受限性、拓扑动态性、协议异构性及攻击多样性等方面存在显著差异,传统IDS难以直接适配。因此,近年来研究者们基于深度学习、联邦学习、区块链及混合检测框架等策略,针对上述场景提出了多种优化方案。
车联网场景下的入侵检测系统应用聚焦于“实时性-混合检测-小样本鲁棒”三条主线,以解决车载控制器算力受限、DoS 攻击多样及训练数据稀缺等痛点。郭健忠等[73]提出树结构堆叠集成与无监督异常检测协同框架,通过贝叶斯优化超参数,在 Car-Hacking 数据集上对已知 DoS 攻击实现 100% 准确率;随后将集成模型标记为“正常”的流量输入 CDAE-AM 模块,针对未知零日攻击召回率达到 100%,整体测试延迟仅 0.072 ms,满足车载 10 ms 实时约束。银鹰等[74]通过分析车载CAN面临的信息安全威胁,提取CAN报文在报文ID,时间间隔,数据字段中的通信特征,提出一种基于长短期记忆(long short term memory,LSTM)的CAN入侵检测模型,该模型能有效保留CAN报文的时序特征,在CAN遭受攻击时检测攻击行为以及对应的攻击类型.实验结果表明,该模型的攻击检测精度达99.99%。为缓解 CAN 报文类别不平衡与样本稀缺问题,王浩轩等[75]提出GAN-Transformer框架,通过掩码预测生成器扩充训练集,判别器完成异常分类,最终在Car-Hacking 数据集上F1达99.66%,检测耗时仅为CNN的4.2%,验证了小样本场景下的高鲁棒性。上述工作共同表明,车联网 IDS 正由单一监督模型演化为“已知-未知攻击协同检测、轻量化推理、数据增强”一体化的实时防御体系。
LiShuai. Report on cybersecurity vulnerability landscape in the first half of 2024[M]. Shantou: The Fifth Electronics Research Institute of the Ministry of Industry and Information Technology, 2024.
[3]
AndersonJ P. Computer security threat monitoring and surveillance: technical report[M]. Fort Washington: James P Anderson Company, 1980.
[4]
DennningD E, NeumannP G. Requirements and model for IDES: a real-time intrusion detection system[M]. Menlo Park: Technical report, Computer Science Laboratory, SRI International, 1985.
[5]
HintonG E, OsinderoS, TehY W. A fast learning algorithm for deep belief nets[J]. Neural Computation, 2006, 18(7): 1527-1554.
DingShunying. Research on defense mode of big data network security based on deep learning[J]. Information and Computer(Theoretical Edition), 2018(17): 194-195.
[8]
BarachJ. Enhancing intrusion detection with CNN attention using NSL-KDD dataset[C]// Artificial Intelligence for Business(AIxB). Laguna Hills: IEEE, 2024: 15-20.
[9]
KasongoS M. A deep learning technique for intrusion detection system using a recurrent neural networks based framework[J]. Compu-ter Communications, 2023, 199: 113-125.
[10]
KostageK, WestD, MeinertT, et al. Enhancing autonomous intrusion detection system with generative adversarial networks[C]// IEEE 20th International Conference on e-Science. Osaka: IEEE, 2024: 1-10.
[11]
LaghrissiF, DouziS, DouziK, et al. Intrusion detection systems using long short-term memory(LSTM)[J]. Journal of Big Data, 2021, 8(1): 65-75.
[12]
WangY, HanZ, DuY, et al. BS-GAT: a network intrusion detection system based on graph neural network for edge computing[J]. Cybersecurity, 2025, 8(1). DOI: 10.1186/s42400-024-00296-8.
[13]
QuyenN H, DuyP T, NguyenN T, et al. FedKD-IDS: a robust intrusion detection system using knowledge distillation-based semi-supervised federated learning and anti-poisoning attack mechanism[J]. Information Fusion, 2025(117). DOI: 10.1016/j.inffus.2024.102807.
[14]
Le JeuneL, GoedemeT, MentensN. Machine learning for misuse-based network intrusion detection: overview, unified evaluation and feature choice comparison framework[J]. IEEE Access, 2021, 9: 63995-64015.
[15]
HajjS, El SibaiR, Bou AbdoJ, et al. Anomaly-based intrusion detection systems: the requirements, methods, measurements, anddatasets[J]. Transactions on Emerging Telecommunications Technologies, 2021, 32(4). DOI: 10.1002/ett.4240.
[16]
OuY J, LinY, ZhangY, et al. The design and implementation of host-based intrusion detection system[C]// Third International Symposium on Intelligent Information Technology and Security Informatics. Ji’an: IEEE, 2010: 595-598.
[17]
ThakkarA, LohiyaR. A survey on intrusion detection system: feature selection, model, performance measures, application perspective, challenges, and future research directions[J]. Artificial Intelligence Review, 2022, 55(1): 453-563.
[18]
AyyagariM R, KesswaniN, KumarM, et al. Intrusion detection techniques in network environment: a systematic review[J]. Wireless Networks, 2021, 27(2): 1269-1285.
[19]
TavallaeeM, BagheriE, LuW, et al. A detailed analysis of the KDD CUP 99 data set[C]// IEEE Symposium on Computational Intelligence for Security and Defense Applications. Ottawa: IEEE, 2009: 1-6.
[20]
ProtićD. Review of KDD Cup’99, NSL-KDD and Kyoto 2006+datasets[J]. Vojnotehnicki Glasnik, 2018, 66(3): 580-596.
[21]
SongJ, TakakuraH, OkabeY, et al. Statistical analysis of honeypot data and building of Kyoto 2006+ dataset for NIDS evaluation[C]// Proceedings of the First Workshop on Building Analysis Datasets and Gathering Experience Returns for Security. Salzburg: ACM, 2011: 29-36.
[22]
Mohd YusofN, SulaimanN S. Cyber attack detection dataset: a review[J]. Journal of Physics: Conference Series, 2022, 2319(1). DOI: 10.1088/1742-6596/2319/1/012029.
[23]
MoustafaN, SlayJ. UNSW-NB15:a comprehensive data set for network intrusion detection systems(UNSW-NB15 network data set)[C]// Military Communications and Information Systems Conference. Canberra: IEEE, 2015: 1-6.
[24]
SharafaldinI, Habibi LashkariA, GhorbaniA A. Toward generating a new intrusion detection dataset and intrusion traffic characterization[C]// Proceedings of the 4th International Conference on Information Systems Security and Privacy, Funchal. Madeira: SCITEPRESS-Science and Technology Publications, 2018: 108-116.
[25]
SahebM C P, YadavM S, BabuS, et al. A review of DDoS evaluation dataset:CICDDoS 2019 dataset[C]// Energy Systems, Drives and Automations, ESDA 2021.Lecture Notes in Electrical Engineering. Singapore: Springer, 2021: 1-11.
[26]
CreechG, HuJ. A semantic approach to host-based intrusion detection systems using contiguous and discontiguous system callpatterns[J]. IEEE Transactions on Computers, 2014, 63(4): 807-819.
[27]
SunY. Modifying the one-hot encoding technique can enhance the adversarial robustness of the visual model for symbol recognition[J]. Expert Systems with Applications, 2024(250). DOI: 10.1016/j.eswa.2024.123751.
[28]
GoodfellowI, BengioY, CourvilleA. Deeplearning[M]. Cambridge: MIT Press, 2016.
[29]
GéronA. Handson machine learning with scikit-learn, keras, and tensorflow[M]. Sebastopol: O’reilly Media, 2019.
TangHaoyao, CuiXin, ZhangYiwei, et al. Network traffic classification method improved based on data augmentation and CNN-Optuna-Attention[J]. Science Technology and Engineering, 2025, 25(15): 6419-6430.
[33]
BishopC M. Pattern recognition and machine learning[M]. Berlin: Springer, 2006.
[34]
LeeW, StolfoS J. Data mining approaches for intrusion detection[C]// USENIX Security Symposium. New York: USENIX Associa-tion, 1998. DOI: 10.1109/45.624335.
[35]
Ozkan-OkayM, SametR, AslanO, et al. A comprehensive systematic literature review on intrusion detection systems[J]. IEEE Access, 2021, 9: 157727-157760.
[36]
HintonG E, SalakhutdinovR R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786): 504-507.
[37]
AbdiH, WilliamsL J. Principal component analysis[J]. Wiley Interdisciplinary Reviews: Computational Statistics, 2010, 2(4): 433-459.
[38]
WangH, RajB. On the origin of deep learning[J]. arXiv Preprint, arXiv: 1706.03762, 2017.
[39]
LecunY, BoserB, DenkerJ S, et al. Backpropagation applied to handwritten zip code recognition[J]. Neural Computation, 1989, 1(4): 541-551.
[40]
YamashitaR, NishioM, DoR K G, et al. Convolutional neural networks: an overview and application inradiology[J]. Insights into Imaging, 2018, 9(4): 611-629.
[41]
YangH, WangF. Wireless network intrusion detection based on improved convolutional neural network[J]. IEEE Access, 2019, 7: 64366-64374.
[42]
SamhaA K, MalikN, SharmaD, et al. Intrusion detection system using hybrid convolutional neural network[J]. Mobile Networks and Applications, 2023. DOI: 10.1007/s11036-023-02223-6.
[43]
PengT E, LiuI H, LiJ S, et al. Network intrusion detection using CNN-based classification method[C]// IEEE/ACIS 8th International Conference on Big Data, Cloud Computing, and Data Science. Vietnam: IEEE, 2023: 361-364.
[44]
XueY, KangC, YuH, HAE-HRL: a network intrusion detection system utilizing a novel autoencoder and a hybrid enhanced LSTM-CNN-based residual network[J]. Computers & Security, 2025, 151. DOI: 10.1016/j.cose.2025.104328.
[45]
ElmanJ L. Finding structure intime[J]. Cognitive Science, 1990, 14(2): 179-211.
[46]
BergstraJ, BreuleuxO, BastienF, et al. Theano: a CPU and GPU math compiler in python[C]// Python in Science Confe-rence. Austin: IEEE, 2010: 18-24.
[47]
BahdanauD, ChoK, BengioY. Neural machine translation by jointly learning to align and translate[J]. arXiv Preprint, arXiv: 1706.037, 2016.
[48]
GravesA, MohamedA R, HintonG. Speech recognition with deep recurrent neural networks[C]// IEEE International Confe-rence on Acoustics, Speech and Signal Processing. Vancouver: IEEE, 2013: 6645-6649.
[49]
HochreiterS, SchmidhuberJ. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780.
[50]
LecunY, BottouL, BengioY, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.
[51]
DuX, DingX, TaoF. Network security situation prediction based on optimized clock-cycle recurrent neural network for sensor-enabled networks[J]. Sensors, 2023, 23(13). DOI: 10.3390/s23136087.
[52]
AzarudeenK, GhulamD, RakeshG, et al. Intrusion detection system using machine learning by RNN method[J]. E3s Web of Conferences, 2024, 491. DOI: 10.1051/e3sconf/202449104012.
[53]
ChavhanN. Statistical implementation for SD-RNN model for multi-class classification for network intrusion detection system[J]. Communications on Applied Nonlinear Analysis, 2024, 31(2S): 546-560.
[54]
GoodfellowI, Pouget-AbadieJ, MirzaM, et al. Generative adversarial nets[C]// Proceedings of the 28th International Conference on Neural Information Processing Systems. Cambridge: MIT Press, 2019: 2672-2680.
[55]
KarrasT, LaineS, AilaT. A style-based generator architecture for generative adversarial networks[C]// IEEE/CYF Conference on Computer Vision and Pattern Recognition(CVPR). Long beach: IEEE, 2019: 4396-4405.
[56]
DentonE L, ChintalaS. Deep generative image models using a laplacian pyramid of adversarial networks[C]// Proceedings of the 29th International Conference on Neural Information Processing Systems. Cambridge: MIT Press, 2018: 1486-1494.
[57]
ShinH C, TenenholtzN A, RogersJ K, et al. Medical image synthesis for data augmentation and anonymization using generative adversarial networks[M]. Berlin: Springer, 2018.
LeeJ, ParkK. Gan-based imbalanced data intrusion detectionsystem[J]. Personal and Ubiquitous Computing, 2021, 25(1): 121-128.
[60]
ShahriarM H, HaqueN I, RahmanM A, et al. G-ids: generative adversarial networks assisted intrusion detection system[C]// IEEE 44th Annual Computers, Software, and Applications Confe-rence. Madrid: IEEE, 2020: 376-385.
[61]
LiaoD, HuangS, TanY, et al. Network intrusion detection method based on GAN model[C]// International Conference on Computer Communication and Network Security(CCNS). Xi’an: IEEE, 2020: 153-156.
[62]
ArafahM, PhillipsI, AdnaneA, et al. An enhanced BiGAN architecture for network intrusion detection[J]. Knowledge-Based Systems, 2025, 314. DOI: 10.1016/j.knosys.2025.113178.
[63]
HintonG E. Learning distributed representations of concepts[C]// Proceedings of the 8th Annual Conference of the Cognitive Science Society. Amherst: Cognitive Science Society, 1986: 1-12.
[64]
VincentP, LarochelleH, BengioY, et al. Extracting and composing robust features with denoisingautoencoders[C]// Proceedings of the 25th International Conference on Machine Learning-ICML. Helsinki: ACM Press, 2008: 1096-1103.
[65]
MasciJ, MeierU, CiresanD, et al. Stacked convolutional auto-encoders for hierarchical feature extraction[M]. Verlag: Springer, 2011.
AndresiniG, AppiceA, MalerbaD. Autoencoder-based deep metric learning for network intrusiondetection[J]. Information Sciences, 2021, 569: 706-727.
[68]
YashwanthT, AshwiniK, ChaithanyaG S, et al. Network intrusion detection using auto-encoder neural networks and MLP[C]// Third International Conference on Distributed Computing and Electrical Circuits and Electronics. Ballari: IEEE, 2024: 1-6.
[69]
YangY, ZhengK, WuC, et al. Building an effective intrusion detection system using the modified density peak clustering algorithm and deep belief networks[J]. Applied Sciences, 2019, 9(2). DOI: 10.3390/app9020238.
[70]
ZhangH, LiY, LüZ, et al. A real-time and ubiquitous network attack detection based on deep belief network and support vectormachine[J]. IEEE/CAA Journal of Automatica Sinica, 2020, 7(3): 790-799.
[71]
ChenA, FuY, ZhengX, et al. An efficient network behavior anomaly detection using a hybrid DBN-LSTM network[J]. Compu-ters & Security, 2022, 114. DOI: 10.1016/j.cose.2021.102600.
[72]
CaiZ, DaiY, ZhangJ, et al. SA-ResNet: an intrusion detection method based on spatial attention mechanism and residual neural network fusion[J]. Computers, Materials & Continua, 2025, 83(2). DOI: 10.32604/cmc.2025.061206.
CaiLizhi, WangXiaoyu, YangKang, et al. An IoT intrusion detection system based on multiple feature fusion[J]. Computer Technology and Development, 2025, 35(9): 85-92.
GuoJianzhong, WangCan, XieBin, et al. Hybrid intrusion detection system for DoS attacks in internet of vehicles[J]. Compu-ter Systems & Applications, 2025, 34(3): 85-93.
FuYu, WangKun, DuanXueyuan, et al. Survey of research on abnormal traffic detection for software defined networks[J]. Journal on Communications, 2024, 45(3): 208-226.
[87]
TayfourO E, MubarakaliA, TayfourA E, et al. Adapting deep learning-LSTM method using optimized dataset in SDN controller for secure IoT[J]. Soft Computing, 2023, 5: 1-9.
[88]
LongZ, JinsongW. A hybrid method of entropy and SSAE-SVM based DDoS detection and mitigation mechanism in SDN[J]. Computers & Security, 2022, 115. DOI: 10.1016/j.cose.2022.102604.