A Study on Convolutional Neuron Networks for Detecting Network Intrusions Using Deep Learning Approach
碩士 === 崑山科技大學 === 資訊管理研究所 === 106 === Most existing approaches for analysing network threats uses machine learning approaches to discriminate the behaviour differences between normal and malicious connections by collecting a large number of network connection packets. Generally, it requires a great...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | zh-TW |
| Published: |
2018
|
| Online Access: | http://ndltd.ncl.edu.tw/handle/4h4eyc |
| Summary: | 碩士 === 崑山科技大學 === 資訊管理研究所 === 106 === Most existing approaches for analysing network threats uses machine learning approaches to discriminate the behaviour differences between normal and malicious connections by collecting a large number of network connection packets. Generally, it requires a great deal of manpower and resources and cannot cope with the new network threats with diverse signatures, so malicious network intrusion detection needs a quick and precise approach to detect the network intrusions. Defenders found that there are many similar behavioral features of different threats using statistical analysis of threat patterns. Deep Learning (DL) is essentially a multi-layer deep neural network (DNN) architecture that learns common behavioral features to categorize new network threats. The present study developed an improved behaviour-based classifier learning model for DDoS detection by training an CNNs (Convolutional Neural Networks) with TensorFlow developed by Google to extract the behaviour features from network flows to form the feature matrix (bitmap format). The study revised the architecture of original LeNet-5 model for learning of network behavioral features to detect malicious network threats. Finally, identify the class of threats with Softmax function by using the optimal weights of hidden layers with error correction between the estimated value and the actual output using optimization algorithm. The experimental results show that our approach cannot only increase the learning speed of network behavioural pattern, but also improve the accuracy of network intrusion detection to reduce the threat of network attacks.
|
|---|