Synthesis of a Machine Learning Model for Detecting Computer Attacks Based on the CICIDS2017 Dataset

The paper deals with the construction and practical implementation of the model of computer attack detection based on machine learning methods. Among available public datasets one of the most relevant was chosen - CICIDS2017. For this dataset, the procedures of data preprocessing and sampling were developed in detail. In order to reduce computation time, the only class of computer attacks (brute force, XSS, SQL injection) was left in the training set. The procedure of feature space construction is described sequentially, which allowed to significantly reduce its dimensions - from 85 to 10 most important features. The quality assessment of ten most common machine learning models on the obtained pre-processed dataset was made. Among the models (algorithms) that demonstrated the best results (k-nearest neighbors, decision tree, random forest, AdaBoost, logistic regression), taking into account the minimum time of execution, the choice of random forest model was justified. А quasi-optimal selection of hyper parameters was carried out, which made it possible to improve the quality of the model in comparison with the previously published research results. The synthesized model of attack detection was tested on real network traffic. The model has shown its validity only under the condition of training on data collected in a specific network, since important features depend on the physical structure of the network and the settings of the equipment used. The conclusion was made that it is possible to use machine learning methods to detect computer attacks taking into account these limitations.

PDF Abstract


Results from the Paper

Ranked #2 on Network Intrusion Detection on CICIDS2017 (using extra training data)

     Get a GitHub badge
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Network Intrusion Detection CICIDS2017 Random Forest Avg F1 0.971 # 2
Recall 96.1 # 4
Precision 98.2 # 2


No methods listed for this paper. Add relevant methods here