|
An Efficient Autoencoder Based Deep Learning Technique to Detect Network Intrusions |
|---|---|
| รหัสดีโอไอ | |
| Creator | C. Haripriya, M.P. Prabhudev Jagadeesh |
| Title | An Efficient Autoencoder Based Deep Learning Technique to Detect Network Intrusions |
| Contributor | - |
| Publisher | TuEngr Group |
| Publication Year | 2565 |
| Journal Title | International Transaction Journal of Engineering, Management, & Applied Sciences & Technologies |
| Journal Vol. | 13 |
| Journal No. | 7 |
| Page no. | 13A7P: 1-9 |
| Keyword | Intrusion Detection, Deep learning, Autoencoder, CSE_CIC-IDS 2018, Class Imbalance, SMOTE, Cybersecurity |
| URL Website | http://TuEngr.com/Vol13-7.html |
| Website title | ITJEMAST V13(7) 2022 @ TuEngr.com |
| ISSN | 2228-9860 |
| Abstract | With the tremendous advancements in internet technology, the amount of data generated over the network is very large. In a network connected with millions of computers, Terabytes/Zettabytes of data are generated every second. It is almost impossible to analyze this enormous data generated in the network manually. Companies have to incur huge losses if their network is compromised, hence timely detection of intrusions is very important to help the organizations prevent further attacks. Deep learning algorithms proved to be more effective when compared to Machine Learning algorithms. Earlier research works focused on old sets like KDDCup99 which do not reflect current-day attacks. All the Intrusion Detection Datasets are imbalanced and have severely skewed class distribution. Many researchers do not focus on class imbalance and their classification models tend to overfit. The major motivation of our research work is to focus on data pre-processing techniques and address the class imbalance problem using SMOTE (Synthetic Minority Oversampling Technique). We implement a deep autoencoder on the latest dataset which is the latest benchmark dataset that reflects current attacks. The average accuracy considering all the CSV (Comma Separated Values) files of the "CSE-CIC-IDS 2018" is 97.79 %. The proposed model achieved promising results and is more accurate since we considered all the records and attack types of the dataset. |
