Abnormal Traffic Detection In Railway Signal System Network Based On Deep Learning

Railway signal system is a key system to ensure railway traffic safety and improve transportation efficiency.The railway signal system network is a network constructed to realize the data transmission between the various subsystems of the railway signal system and to exchange information with other railway systems.Therefore,the safety of the signal system network will directly affect the operation safety and operation efficiency of the train.According to the traditional concept,the railway signal system network has a certain closedness.It has always been considered a non-invasive risk.With the rapid development of computer network technology and the continuous improvement of service requirements,the railway signal system will inevitably interface with other railway information systems.These railway information systems have strong public service features.They inevitably have access to open networks such as the Internet.Therefore,they inevitably provide an opportunity for attackers to invade.Moreover,this may cause the paralysis of the signal system and the loss of information transmission errors.To reduce the possibility of intrusion into the railway signal system network,it is necessary to introduce intrusion detection technology into the network.As an important means of network intrusion detection,abnormal traffic detection has become the main method to prevent network intrusion.At present,detection methods based on deep learning are popular because they can automatically train and analyze the multi-layer features of attack traffic and use them for classification.In the abnormal traffic detection of railway signal system network,there are problems such as weak generalization ability,unbalanced sample distribution,poor feature selectivity,poor few-shot sample detection performance,poor anti-noise performance and efficiency,and poor special attack detection ability.Therefore,designing a detection model with better performance is still of great significance to the security of railway signal system networks based on deep learning theory.After studying the features of abnormal traffic of railway signal system network and various deep learning techniques,the abnormal traffic detection of railway signal system network in different specific scenarios is completed according to the above problems in this dissertation.The main work is as follows:(1)Aiming at the problems of weak generalization ability and unbalanced distribution of traffic samples in abnormal traffic detection of railway signal system network,a method based on feature extraction and IDELM is proposed.The procedure is comprised of two sections: To begin,a method for detecting abnormal traffic is proposed using an optimized long short memory network(LSTM)and an improved residual neural network.First,a three-layer stacked LSTM network is used to extract features of network traffic at various depths.Then,Inception modules are combined to build the data pooling layer.Finally,an improved residual neural network with skip connection line is designed to improve the accuracy of abnormal traffic detection in the network.In the second section,an unbalanced abnormal traffic detection method based on improved Res-BIGRU and integrated dynamic ELM optimization is proposed to address the problem of unbalanced distribution of traffic samples.First and foremost,an improved Res-BIGRU structure based on the GRU network is proposed.Consequently,the traffic’s feature extraction effect is further enhanced.The strategy utilizes two IDELMs to update the final classification results,thereby improving the final detection accuracy.Experiments demonstrate that the overall detection performance is superior to that of current intrusion detection methods.It is extremely robust that data samples are damaged.Additionally,it possesses an excellent capacity for generalization.When using the improved Res-BIGRU and IDELM structure to detect unbalanced traffic samples,it can achieve higher detection accuracy and better robustness in the railway signal system network environment.(2)Aiming at the problem of poor model feature selection and few-shot sample abnormal traffic detection in railway signal system network,an abnormal traffic detection method based on improved GAN and feature optimization is proposed.Additionally,the method is divided into two sections: To begin,a method for detecting abnormal traffic is proposed based on Generative Adversarial Networks(GANs)and feature optimization.The method utilizes a collaborative automatic learning machine strategy to optimize feature selection.Meanwhile,an MMD-GAN network is proposed using GAN and multi-kernel mean difference to mix random noise data and original training label samples,and the detection accuracy is improved.Aiming at the attitude transformation and few-shot sample traffic detection of the railway signal system,a method abnormal traffic using multi-scale Deep-CapsNet and adversarial reconstruction is proposed.Iterative routing is optimized by enhancing the EM vector clustering of Deep-CapsNet.Additionally,Deep-CapsNet is optimized using a multi-scale convolutional capsule.Finally,an approach is written for an adversarial reconstruction classification network(ARCN).It is accomplished through the use of supervised source data classification and unsupervised target domain data reconstruction.Adversarial training is combined with the reconstruction of few-shot traffic features to reduce noise interference.The experimental results show that the proposed method has good feature selection performance.In the few-shot sample traffic detection of railway signal system network,the detection performance is the best.It can be used for real-time few-shot sample abnormal traffic detection.(3)To address the noise interference problem in train control system network attack traffic detection,a method based on channel boosting convolutional neural networks(CB-CNNs)and segmentation residual network optimization is proposed.To begin,normal traffic is modeled using a denoising auto-encoder(DAE).Then,channel boosting is proposed to convert the correlation error vector to the multi-channel input of the classifier.Finally,a residual network with multi-path segmentation is designed to optimize CB-CNN.Thereby,the abnormal traffic detection performance in the railway signal system network is optimized.The experimental results indicate that the proposed method performs better during training.When the false positive rate is low,it has a favorable classification visualization effect.Additionally,real-time detection outperforms other detection methods.It can be used to train control networks to detect attack traffic in complex scenes.Moreover,the robustness is improved in the presence of other sample interferences.(4)In view of the strong concealment of LDoS attacks in the railway signal system network,traditional methods based on signal analysis are difficult to detect the lack of LDoS attack traffic in the normal traffic with large fluctuations.In this dissertation,an LDoS attack detection method is proposed based on Laplacian feature map and DP-WGAN.It is constructed using Laplacian Eigenmaps(LE)and Wasserstein adversarial generative networks with a denoising penalty(DP-WGAN).To begin,an unsupervised deep learning network(LENet)is created using LE.Moreover,the data is preprocessed using dimensionality reduction to extract lowdimensional LDoS features.Then,a DP-WGAN discriminant network on top of WGAN is built to detect LDoS attacks.Moreover,a denoising penalty term and a sample augmented discriminator are created to improve the LDoS attacks features generated.Denoising penalty is employed to improve the generator’s performance.Moreover,the sample augmented discriminator learns a more accurate classification decision hyperplane.As a result,LDoS attacks detection is more precise.The results show that the proposed method has high detection accuracy and efficiency on the two datasets.It is also highly robust when the samples contain noise.Moreover,LDoS attacks can be detected in real time.