Backdoor Attacks Prediction in IIoT Network using Optimal Double Mask Region Convolution Model

Abstract

The evolution of Industrial Internet of Things (IIoT) applications provides intelligent microservices to process the generated massive data. However, the open and interconnected structure of IIoT makes the network model more vulnerable to malware attacks. Among different kinds of threats, the backdoor attack is considered the main imperceptible attack due to its unobservable and opaque characteristics. Therefore, to accurately predict backdoor attacks in the network model and to mitigate them efficiently, this paper proposes a novel concept named “novel binary knowledge gaining based double mask region convolution (NBG-DRC) model”. The main objective of the proposed concept is two-fold: backdoor attack prediction and trigger identification. The backdoor attack prediction module uses a double mask region convolution (DRC) network to predict whether the data is malicious or normal. Subsequently, in the trigger identification phase, the structure and positions of backdoor triggers are identified accurately using the NBG-DRC model. The proposed model uses four different datasets namely CIFAR-10, MNIST, CIFAR-100, and GTSRB for analysis. The efficiency of the proposed NBG-DRC model is examined by comparing its performance rate with other compared techniques in terms of evaluation metrics namely false negative rate, false positive rate, computational overhead, execution time, accuracy, and success rate. The proposed NBG-DRC technique achieves a greater accuracy percentage of about 94% for the CIFAR-10 dataset, 98.7% for the MNIST dataset, 91% for the CIFAR-100 dataset, and 97.23% for the GTSRB dataset.