A Novel Fuzzing for RPL Network Vulnerability Analysis and Vision Transformer-based Attack Detection for IIoT

Abstract

Recent scientific advancements in information and communication technologies enable Industry 4.0 (I4.0), which empowers smart manufacturing with unprecedented operational efficiency and productivity. Integrating the smart Industrial Internet of Things (IIoT) facilitates continuous real-time monitoring of manufacturing processes by establishing safety controls through data collection. Despite the substantial benefits, the vast network of interconnected IoT devices in the I4.0 environment is vulnerable to cyber security threats. Routing Protocol for Low Power Lossy Networks (RPL) is a reliable, energy-efficient, and flexible way to set up a routing framework for IIoT-based critical industrial communication infrastructure. However, network security is a critical concern in RPL-based IIoT environments due to complex patterns and subtle deviations in the behavior of the network. Therefore, it is crucial to introduce novel security solutions with more accurate vulnerability analysis and attack detection. This work proposes a Novel RPL Security (NRS) approach that includes fuzzing-based vulnerability analysis and vision transformer-based attack detection to solve the abovementioned issues. The proposed work encompasses two primary components: the Wasserstein Generative Adversarial Network (WGAN)-based fuzzing method for RPL network vulnerability analysis and vision transformer-based attack discovery. In the first method, routing data from the RPL-IIoT network is collected, and the fuzzing model is combined with the WGAN to improve the vulnerability distribution in the fuzzer output data. The analyzed fuzzer output data is converted into images and fed into the vision transformer model for attack discovery. The vision transformer improves attack detection accuracy by effectively capturing complex patterns and subtle deviations in network behavior. Moreover, the efficacy of the proposed NRS is evaluated using Contiki/Cooja-based simulations and Python-based machine-learning models. The results are validated for vulnerability analysis and attack detection using various metrics such as detection accuracy, fuzzer output recognition rate, triggered efficiency of vulnerabilities, and diversity of generated data, revealing the notable outcome of the proposed approach.