Deep Learning for Cyber Attacks Detection in IoT Networks

Abstract

The rapid expansion of Internet of Things (IoT) devices introduces new cybersecurity challenges. Their limited computational power and basic security mechanisms expose them to various attacks such as Distributed Denial of Service (DDoS), spoofing, bruteforce, and reconnaissance. This work investigates deep learning techniques for detecting malicious behavior in IoT networks. We use the CICIoT2023 dataset to simulate IoT traffic and known attacks, and evaluate three neural network architectures: Convolutional Neural Network (CNN) (for spatial features), Long Short-Term Memory (LSTM) (for temporal patterns), and a hybrid CNN-LSTM model. To address class imbalance in the dataset, we apply the SMOTE (Synthetic Minority Oversampling Technique) to generate synthetic samples for underrepresented attacks. Experimental results show that the CNN-LSTM model achieved the best performance with 98.51% accuracy and a 93% F1-score without SMOTE. After applying SMOTE, CNN and LSTM achieved 97.65% and 97.56% accuracy respectively. The hybrid model proved to be the most robust, leveraging spatial and temporal features for effective multi-stage attack detection.