Towards Attention Mechanism in Detecting Botnet Attacks for loT Devices

The widespread adoption of Internet of Things (loT) devices has resulted in great convenience and efficiency. However, it has also initiated new cybersecurity threats like botnet attacks, which lead to a significant challenge. Detecting and reducing these threats in the loT landscape requires innovative approaches that contemplate the unique traits of loT devices. This paper uses Convolutional Neural Networks (CNNs) with Squeeze-and-Excitation (SE) Blocks to detect botnet attacks. We have modified the SE block to apply feature-based attention and introduce weight adjustment, enabling us to gain insight into feature importance. We employ the model on the two subsets, Full Subset and 10-Best features Subset, of the BoT-loT datasets for detecting three major attacks: DDoS, DoS, and Reconnaissance. The overall detection accuracy of our model is 99.96%. We extend our experiments by applying the model for protocol-based botnet attack detection. Our proposed methods perform significantly well for the UDP protocol for both subsets, achieving 99.99% accuracy. Regarding the TCP protocol, the model attains 99.92% accuracy for the Full Subset and 97.55% for the 10-Best Features subset. In addition, we perform a comparative analysis with other Deep Learning models and ML Classifiers. Although Stacked LSTM performed similarly well for UDP, our proposed methods surpassed other experimented models in terms of TCP.