GSSB: A NOVEL METHOD FOR PROVIDING A HYBRID DEEP LEARNING MODEL FOR ANOMALY DETECTION IN INTERNET OF THINGS NETWORKS

Abstract

In recent years, extensive research has been published to detect anomaly in the Internet of Things, but this research has not been able to respond to challenges such as limited coverage of attack types, data imbalance, lack of effective optimization mechanisms, and inefficiency of sensor nodes for deep learning, etc. This paper presents an intelligent and decentralized intrusion detection system for IoT networks using a combination of deep learning techniques, artificial data generation, and optimization algorithms. The main objective is to design a model called GSSB (GAN_SCNN_SHO_BiLSTM) that is able to accurately detect various cyber-attacks, especially unknown and zero-day attacks. In this method, to slove the problem of imbalance data from Generative Adversarial Networks (GAN) are used, after generating augmented data by GAN, the Hurst profile is extracted as a complementary statistical feature by calculating the autocorrelation and stability of the data and added to the feature vector to increase the model's ability to recognize spatial patterns and temporal dynamics. Then complex location features of network traffic with the Stacked Convolutional Neural Network (SCNN) are extracted and the time data analysis was done by the BiLSTM network. Also used to select optimal features and reduce model complexity of the model, the binary version of the Seahorse Optimization (SHO) algorithm is used. The proposed structure is designed in a Fog Computing environment and uses federated learning to preserve data privacy and models trained in sub-nodes are aggregated together. To evaluate the model, experiments were conducted on the dataset Edge-IIoT 2023 and Edge-IIoT, and performance of the model was compared with criteria such as Accuracy, Recall, Detection Rate, False Positive Rate (FPR), and F1 Score. The results showed that the proposed model managed to achieve a F1 score of 99.3%, a recall of 98.6%, and an accuracy of 99.2%, and performed better compared to other reference methods such as CNN-BiLSTM and the federated method. These results demonstrate the ability of the GSSB model to identify advanced threats and improve the security of IoT networks in real-time situations.