The Industrial Internet of Things (IIoT) has emerged as a prominent area of
research, leveraging sensors, actuators, and computing capabilities to revolutionize data
collection, communication, and processing. This has led to the development of
innovative Industry 4.0 applications across various sectors such as mining, energy,
healthcare, agriculture, and transportation. Despite its numerous benefits, security and
privacy remain significant challenges in IIoT design, necessitating the integration of
intrusion detection systems (IDS) for mitigation. Recent advancements in machine
learning (ML) and deep learning (DL) algorithms offer promising avenues for the
development of effective IDS techniques tailored to the IIoT environment. However,
the scarcity and imbalance of attack datasets pose obstacles to existing IDS models. In
response, this paper proposes a deep learning-enabled privacy-preserving technique
aimed at improving classification accuracy. The proposed Fused IGAN-IDS model
comprises three key components: (i) Generative Adversarial Networks (GAN) for
generating high-quality samples to address data imbalance; (ii) An extremely
randomized tree-based feature selection algorithm for identifying significant features
from the dataset; and (iii) The Sparse Stacked Autoencoder (SSAE) algorithm for
selecting relevant features to optimize the attack detection process's timing efficiency.
By incorporating adversarial data samples into the original dataset, the fused IGANIDS model reached an accuracy of 99.6%, sensitivity of 99.5%, specificity of 99.7%,
precision of 98.4%, and a Jaccard Similarity Score (JSS) of 99.7%. This approach also
resulted in a lower false alarm rate, with a 4% False Positive Rate (FPR) and a 3%
False Negative Rate (FNR).
Keywords: AEDLN, BSA, Deep learning, Fused IGAN, Feature selection, GAN, IGAN, Randomized tree, Sparse stacked auto encoder.
