Feature Constraint-based Adversarial Examples Generation Research On Intrusion Detection Classifiers

In recent years,with the development of machine learning algorithms and their wide applications in various fields,the security problems of machine learning models are gradually emerging.By generating adversarial examples which are slightly modified by the real dataset,attackers can lead machine learning models to a false result,which poses a potential security threat for applications of machine learning.Especially in intrusion detection,adversarial examples can evade the detection of intrusion detection systems,resulting in malicious attacks on systems.However,all previous research on the generation of intrusion detection adversarial examples ignores the feature constraints existing in the dataset.Adversarial examples that do not meet the feature constraints are easily detected and cannot attack the system successfully.Therefore,we propose a feature constraint-based intrusion detection adversarial examples generation method.Based on the deep neural network model,we generated the adversarial examples which can satisfy the feature constraints of network traffic dataset.In the domain of intrusion detection,based on the existing research on adversarial examples generation,we focus on the feature constraints of network traffic dataset,and explores the method of generating intrusion detection adversarial examples based on feature constraints.Therefore,we proposed a feature constraint analysis method based on Pearson matrix and an intrusion detection adversarial examples generation method based on C-IFGSM.First of all,we analyze the three characteristics of the network traffic dataset,which are multi-type of features,specific malicious payloads and the feature constraints.Among them,we focused on the analysis of the feature constraints,and defined four feature constraints,namely function constraint,correlation constraint,scope constraint and condition constraint.On the basis of the correlation analysis of Pearson Correlation Coefficient matrix and expert experience analysis,a feature constraint analysis method is proposed.Through the analysis method,four kinds of feature constraints in network traffic datasets can be obtained.Then,based on the analysis results of feature constraints,we propose an intrusion detection adversarial example generation method,which is composed of an intrusion detection model based on Deep Neural Network and an intrusion detection adversarial example generation algorithm based on C-IFGSM.Firstly,the model parameters need be obtained by building the intrusion detection classifier.Then,according to the four feature constraints,the constraint restriction matrix is added to IFGSM algorithm to ensure that the feature constraints remain unchanged when generating adversarial examples.Afterwards,the intrusion detection adversarial examples that meet the feature constraints can be obtained.Finally,we use the NSL-KDD standard dataset to train an intrusion detection classification model,and evaluate the intrusion detection adversarial example generation algorithm proposed in this paper.Based on the four feature constraints in the NSL-KDD dataset,the C-IFGSM algorithm can generate intrusion detection adversarial examples that satisfy the feature constraints successfully.Compared with the adversarial examples generated by FGSM and IFGSM algorithm,we find that the C-IFGSM method can increase the feature constraint satisfaction rate from 0.37 to 1,and reduce the accuracy of intrusion detection classifier from 0.8 to 0.42.