Evaluation of Deep Learning Techniques for Intrusion Detection in Cybersecurity

Doaa El-Shahat, Department of Computer Science, Faculty of Computers and Informatics, Zagazig University, EgyptFollow
Mohamed Mahmoud Abdelaziz, Department of Computer Science, Faculty of Computers and Informatics, Zagazig University, EgyptFollow
Waseem Akram, Graduate School of Engineering Science and Technology, National Yunlin University of Science and Technology, Douliu, TaiwanFollow
Muhammad Jabir Khan, School of Artificial Intelligence and Computer Science, Nantong University, 226019, ChinaFollow

ORCID

Doaa El-Shahat: https://orcid.org/0000-0003-1681-5039

Waseem Akram: https://orcid.org/0000-0003-2734-1844

Muhammad Jabir Khan: https://orcid.org/0000-0002-7983-706X

Keywords

Internet of things, Intrusion systems, Deep learning, Long short term memory, Feature extraction

Article Type

Original Article

Abstract

The increasing prevalence of Internet of Things (IoT) technology is also leading to a rise in potential cybersecurity risks. Therefore, it has become crucial to develop effective solutions using deep learning to enable us to discover the various vulnerabilities that may arise from continuous technological advancements, and to defend against the different types of attacks that fraudsters might use to exploit system weaknesses, such as Distributed Denial of Service (DDoS) attacks, web attacks, and spoofing. Most model failures stem from their reliance solely on detecting spatial patterns and their failure to focus on detecting temporal patterns, resulting in poor performance in detecting temporal patterns that evolve over time, thus making their detection more difficult. This study proposes an advanced framework based on the sliding window technique to collect ten consecutive records, enabling deep learning models to extract hidden time patterns between attacks. This framework was evaluated on five of the existing robust models: AE-LSTM-CNN, CAE-TCN, CNN-LSTM, LSTM, and CNN-BiLSTM. These five robust models were fed onto the CIC-IoT-2023 dataset, addressing the inherent imbalance using a hybrid approach combining under sampling using RandomUnderSampler and SMOTE techniques. The number of features was reduced to minimize noise and scattering using the Random Forest (RF) algorithm. The experimental results showed the superiority of the LSTM model, achieving the highest accuracy at 98.58% and an F1 score of 98.00%. This was followed by the CAE-TCN model, which demonstrated an excellent balance between accuracy (98.56%) and training speed. These results suggest that using sliding window technology with cached deep learning models enhances their ability to detect time-period patterns between attacks, thus facilitating the detection of small-footprint attacks on the network.

How to Cite

El-Shahat, Doaa; Abdelaziz, Mohamed Mahmoud; Akram, Waseem; and Khan, Muhammad Jabir (2026) "Evaluation of Deep Learning Techniques for Intrusion Detection in Cybersecurity," Sustainable Machine Intelligence Journal: Vol. 14: Iss. 3, Article 1.
DOI: https://doi.org/10.63689/3005-3617.1091