The sound of intrusion: A novel network intrusion detection system
我其实就是被这个标题给吸引过来的，通读了一遍文章后，发现其核心Idea就是将网络数据转化为语音信号，再利用深度学习的算法对于语音数据的优势学习能力进行建模，从而提升入侵检测的分类能力。
-
Idea：
就是利用深度学习优势场景（图像和语音）嫁接（网络安全NIDS）。
-
数据集：NSL-KDD和CICIDS2017
-
信号处理sample：
-
ML算法：Logistic regression, Naive Bayes, KNN, Decision tree, Random forest
-
DL算法：CNN, DBN, LSTM, LSTMCNN
-
测试结果（硬核数据）：
1. 从如下数据可以看出，传统的机器学习算法似乎并没有得益于数据的转换。
2. 而深度学习算法在数据转换的前后，优势体现得淋漓尽致。
3. 基于声音的深度学习模型，相比给予原始数据、转换为更多维度数据，以及转换为图像数据，在Accuracy和FAR上都具有一定优势。
4. 除了基于转换为声音数据的深度学习模型binary-classification, multi-classification任务也具备一定竞争力。
-
讨论部分（精彩的总在这里）
-
FAR对于NIDS的设计至关重要，然而很多研究对此存在某种程度的忽略，为此，（作者）干得不错。
2. 深度学习在图像处理和语音识别处理任务的强大能力，为嫁接到NIDS提供了理论和实践的强大基础。
-
结论和展望（未完待续…）
在原基础上对信号处理的优化，评估转换后的特征extracted features，采用更多的深度学习算法，当然也可以尝试更多的数据集。
如果想要测试验证，这里有开源的库可以直接使用：
“An open source libraries have been utilized in our experiments such as [21] and Kaldi [22] which they have been made for environmental Sound Classification using Deep Learning.”
[21] Fing. Environmental sound classification using deep learning. 2019, Accessed: 2019-02-08.
[22] kaldi. Kaldi speech recognition toolkit. 2019, Accessed: 2019-02-08.
参考文献：
Mohammed Y. Aldarwbi, Arash H. Lashkari, Ali A. Ghorbani, The sound of intrusion: A novel network intrusion detection system, Computers and Electrical Engineering,
https://doi.org/10.1016/j.compeleceng.2022.108455.
(https://www.sciencedirect.com/science/article/pii/S004579062200670X)
Abstract: A network intrusion detection system is an essential part of network security research. It detects intrusion behaviors through active defense technology and takes emergency measures such as alerting and terminating intrusions. To this end, with the rapid development of learning technology, various machine-learning-based and deep-learning-based approaches have been developed, but there is a limitation in the detection accuracy. We believe that dealing with network traffic as if they are vibrations, waves, or sounds would allow us to detect intruders better. In this work, we envisioned a novel Network Intrusion Detection System called “the sound of intrusion”. The proposed system transforms the traffic flow features into waves and utilizes advanced audio/speech recognition deep-learning-based techniques to detect intruders. We used several deep-learning-based techniques including long short-term memory, deep belief networks, and convolutional neural networks. The proposed approach has been validated using two well-known and recent benchmark datasets namely NSLKDD and CIC-IDS2017. It achieves the highest detection accuracy, 84.82%, and 99.41%, with the lowest false alarm rate of 0.12% and 0.004% on two common network intrusion detection systems datasets, namely NSL-KDD and CICIDS2017, respectively. It demonstrates improvements over existing approaches, and shows a strong potential for use as a modern Network Intrusion Detection System.
Keywords: Intrusion detection systems; IDS; Anomaly detection; Deep learning; CNN; DBN; LSTM
原创文章，作者：门童靖博士，如若转载，请注明出处：https://www.agent-universe.cn/2022/11/12684.html
