Deep Learning Based Binary and Multi-class Classification Comparison for Anomaly Detection
Abdoulaye Diallo, Lionel Affognon, Chérif Assane Diallo, Eugène Cokou Ezin · 2022 International Conference on Engineering and Emerging Technologies (ICEET) · 2022
Security remains one of the biggest challenges in the IoT field. This is why several machine learning and deep learning techniques are used to set up models capable of monitoring network traffic and detecting security incidents. In this paper, we develop two classifiers for anomaly detection in IoT networks. One of them is binary and the other is multi-class. After the training phase, we test them on different datasets. Firstly, we test them on samples of the same nature as the training data. Secondly, we test models on new data containing unknown attacks. For a good comparison of the two models, we transform the multi-class classifier confusion matrix into a 2 × 2 matrix. Thus, with two confusion matrices having the same shape, we calculate the different metrics and compare the models’ performance. On data with known types of anomalies (attacks), both models give good performance but the multi-class classifier has a slight advantage. classifier. Conversely, on data containing unknown anomalies (attacks), performance drops on both sides. However, this time the binary classifier proves to be more efficient.