| In the era of the Internet of everything,all kinds of IoT devices have become an indispensable part of people’s lives.However,these devices bring convenience to people as well as security issues.Intrusion detection systems play an important role in preventing IoT security threats and protecting them from attacks.With the proliferation of IoT data and the emergence of unknown attacks,intrusion detection based on traditional machine learning cannot fully mine the features of IoT data and result in low accuracy.However deep learning has advantages in feature mining and big data processing.Therefore,based on deep learning,this thesis conducts in-depth research on the intrusion detection techniques.The main work and contributions are as follows:(1)Aiming at improving the accuracy of machine learning in intrusion detection,a D-GRU intrusion detection algorithm and a sliding window method and are proposed.The D-GRU algorithm is based on a recurrent neural network,which fully mines network traffic data with time-series information.The sliding window method is integrated in data preprocessing stage,enhancing the D-GRU’s ability to learn and represent previous data packets.The accuracy of the D-GRU on the UNSW_NB15 dataset is 94.48%for binary classification and 92.45%for multiclassification,21.31%higher than that of machine learning algorithms.Besides,the parameters,training time and prediction time in D-GRU are more advantageous than other deep learning algorithms.(2)Aiming at solving the problem of attack traffic data imbalance in the intrusion detection.Based on the generative adversarial network,the PTWGAN-GP algorithm to synthesize abnormal traffic flows is proposed.The Wasserstein distance is used in the PTWGAN-GP,and a pre-training module is added to the network to reduce the number of iterations during imbalanced data generation and accelerate the speed of convergence.At the same time,for 4 imbalanced data with very few samples in UNSW_NB15,the precision,recall,fl-score is increased by 10.13%,20.25%and 16.06%respectively.The multi-class classification accuracy of D-GRU is increased by 0.75%.The results shows that PTWGAN-GP has a good effect on imbalanced data processing.(3)Based on D-GRU,an IoT intrusion detection system is constructed,including data acquisition,data processing,data storage,device identification,intrusion detection,access control and visualization module.The test result shows that each module functions well and can satisfy the requirements of intrusion detection in IoT. |
