IOT Attack Detection Using Hybrid Deep Learning with Attention Mechanism

IOT Attack Detection Using Hybrid Deep Learning with Attention Mechanism

Mr. A. D. Talole1,  Riya Mahesh Wani2,  Pooja Ravindra Kankhar3,

Pradnya Rahul Gangurde4, Shreya Sharad Nile5

1.Lecturer, Computer Technology, K. K. Wagh Polytechnic, Nashik

2.Student, Computer Technology, K. K. Wagh Polytechnic, Nashik

3.Student, Computer Technology, K. K. Wagh Polytechnic, Nashik

4.Student, Computer Technology, K. K. Wagh Polytechnic, Nashik

5.Student, Computer Technology, K. K. Wagh Polytechnic, Nashik

Abstract - The rapid proliferation of Internet of Things (IoT) networks has transformed modern society by enabling smart cities, healthcare systems, industrial automation, and intelligent homes through large-scale connectivity and real-time data exchange. However, this unprecedented growth has also widened the attack surface, making IoT devices highly vulnerable to cyberattacks such as denial of service, probing, information theft, and other advanced threats that can compromise device integrity, disrupt services, and expose sensitive information. These challenges are further compounded by the inherent limitations of IoT devices, including constrained computational resources, lack of standardized security protocols, and the diversity of communication technologies, which make them difficult to secure using conventional techniques. Traditional intrusion detection systems (IDS) and classical machine learning approaches have shown limited effectiveness in addressing these issues, as they often fail to capture complex traffic patterns, adapt to dynamic environments, and maintain low false positive rates in real-world conditions. To overcome these shortcomings, this project proposes a hybrid deep learning-based intrusion detection framework that integrates multiple complementary neural network architectures to enhance detection accuracy, scalability, and adaptability. In the proposed framework, Convolutional Neural Networks (CNN) are employed for spatial feature extraction from network traffic, Logistic Regression for balancing the Imbalanced IOT data. Furthermore, an attention mechanism is embedded within the model to dynamically highlight the most relevant features, improving interpretability and ensuring that the system focuses on patterns strongly associated with malicious activity. The framework is trained and validated using the widely recognized UNSW-NB15 dataset, which contains diverse attack scenarios and allows the model to generalize effectively to multiple categories of threats. Experimental evaluations demonstrate that the proposed hybrid architecture achieves superior performance compared to baseline IDS models, consistently delivering high accuracy, precision, recall, and F1-score while significantly reducing false positives. For practical implementation, the system is developed in Jupyter Notebook for model training and deployed using Flask, providing a user friendly web-based interface for real-time intrusion detection. This deployment enables network administrators and security professionals to input traffic data and receive immediate threat assessments, supporting rapid response to potential cyber incidents and reducing overall downtime.

Keywords: CNN (Convolutional Neural Networks), LR (Logistic Regression), Attention Mechanism, UNSW-NB15 dataset, Internet of Things (IoT), Intrusion Detection System (IDS), Threat Intelligence, Hybrid Deep Learning, Cybersecurity.