# Hybrid recurrent with spiking neural network model for enhanced anomaly prediction in IoT networks security

## TL;DR

This paper introduces a hybrid neural network model combining RNN and SNN to improve anomaly detection in IoT networks, achieving high accuracy on benchmark datasets.

## Contribution

A novel hybrid RNN-SNN architecture called HRSNN for enhanced IoT network security with improved anomaly detection.

## Key findings

- HRSNN achieved 99.5% accuracy on the CIC-IoT2023 dataset.
- The model reached 98.75% accuracy on the TON_IoT dataset.
- The hybrid model outperformed existing deep learning approaches in detecting IoT anomalies.

## Abstract

As the number of Internet of Things (IoT) devices grows quickly, cyber threats are becoming more complex and increasingly sophisticated; thus, we need a more robust network security solutions. Traditional deep learning approaches often suffer in identifying effectively anomalies in IoT network. To tackle this evolving challenge, this research proposes a hybrid architecture of Neural Network (NN) models that combine Recurrent-NN (RNN) and Spiking-NN (SNN), referred to as HRSNN, to improve IoT the security.

The proposed HRSNN technique has five steps: preprocessing data, extracting features, equalization classes, features optimization and classification. Data processing step makes sure that input data is accurate and consistent and by employing normalization and the removal of outliers’ techniques. Feature extraction makes use of the RNN part to automatically detect abnormal patterns and high-level features, which are then turned into spike trains for the SNN to process over time. In class equalization step, the Synthetic Minority-Oversampling Technique (SMOTE) is being used resulting in balanced classes. Recursive Feature Elimination (RFE) is used to keep the important features for feature optimization. Then, the dataset is split into sets for testing and training so that the model can be tested properly.

The hybrid model integrates the spatial feature learning skills of RNNs with the temporal adaptability of SNNs, results in an improved accuracy and resilience in identifying IoT network abnormalities. The proposed HRSNN approach, which was tested on the CIC-IoT23 and TON_IoT data sets, achieved better performance compared to current deep learning (DL) models. In particular, experimental assessments show that the model attained an accuracy rate of 99.5% on the “CICIoT2023” dataset and 98.75% on the “TON_IoT” dataset.

These results confirm demonstrate that the proposed architecture of RNN and SSN can achieve significant advancement to IoT security. By combining both spatial and temporal feature learning, HRSNN can improve accuracy detection against diverse security threats. The model is reliable, accurate, and adaptable for safeguarding IoT networks against diverse security threats. Thus, the model addresses the potential solutions in the challenging problem of secured IoT networks.