**1.1 Related work**

Several machine learning, bioinspired, and meta-heuristic methods have been developed for anomaly detection in communication networks. Machine learning algorithms used for intrusion detection can be divided into two broad groups. Deep learning models have achieved remarkable results and can automatically learn feature representations, such as Convolutional Neural Network (CNN) [10], and Deep Neural Network (DNN) [3]. Traditional machine learning techniques, conversely, are characterized for their lack of "depth" in the analysis, such as Support Vector Machine (SVM) [11] K-Nearest Neighbor (KNN) [12], Decision Forest [2], Random Forest [3] and Naive Bayes classifier (NB) [13].

Artificial Immune Systems (AIS) are classified into two major categories, namely network-based and population-based. Network-based algorithms make use of the Immune Network Theory and are based on Artificial Immune Networks [14]. Population-based algorithms, on the other hand, imitate immune cell behavior through artificial agent interactions and are based on Negative Selection [15], Clonal Selection [16, 17], or Danger Theory [8, 18–23]. AIS models have focused on imitating some characteristics of the HIS, such as multiple-level detection mechanisms based on DT [20], and modifications to the DCA. Said modifications include incorporating probability theory [19], fuzzy inference systems [21], feature selection [22], and detection improvements in a semi-supervised context [23].

### **1.2 Contribution**

The main contribution of this research is a biologically inspired NIDS approach based on the deterministic DCA [6]. This model aims to tackle two challenges (and contemporary issues) of NIDS, namely feature selection, and generalization capabilities to improve classification accuracy. A comparison with different bioinspired and machine learning techniques using two publicly available benchmark datasets

(NSL-KDD and UNSW-NB15) is presented. The rest of this paper is organized as follows. Section 2 details the related methodology, as well as the proposed model. Section 3 presents datasets definition, model parameters, and numerical results, as well as a comparison of efficiency metrics with state of the art approaches for binary classification. Section 4 presents conclusions, challenges, and future work.
