Journal:Expert Systems with Applications

Impact Factor:9.4

Summary:As the digital transformation of the industrial environment accelerates, the application of deep learning in the Industrial Internet of Things (IIoT) has become increasingly popular. However, deep neural networks (DNNs) face a serious threat known as “poison attack”, which is particularly prominent in industrial environments. Unlike the image domain, where noticeable perturbations are typically required to influence model behavior, industrial data exhibit high stability, enabling attackers to compromise model integrity using only small perturbations. This “low-magnitude” vulnerability challenges existing poisoning defenses, most of which are adapted from the image domain and thus fail to capture the unique properties and subtle variation patterns of industrial data, resulting in reduced effectiveness in IIoT scenarios. To address these limitations, this paper proposes a poisoning attack mitigation framework tailored for centralized industrial deep learning classification models, covering both the data uploading/storage stage and the model training stage, for which we propose a precise data sanitisation method and a model enhancement method. Experiments conducted on multiple industrial datasets across various application scenarios demonstrate the effectiveness of the proposed approach, outperforming existing defense methods.

Key Words:Poisoning attack; Industrial internet of thing; Deep neural network; Industrial security

Indexed by:SCI JCR Q1

Note:https://www.sciencedirect.com/science/article/abs/pii/S0957417426021536

Discipline:Engineering

Document Type:JCR 一区

First-Level Discipline:Computer Science and Technology

Translation or Not:No