A sentence-level text adversarial attack algorithm against IIoT based smart grid

Abstract

With the development of data processing technologies, efficiency of information processing in the Industrial Internet of Things (IIoT) is greatly improved. In this situation solving the following security problems of the IIoT is the top priority. In IIoT based smart grid, through Natural Language Processing (NLP) technology various types of text data such as the equipment status and historical records can be better utilized. While bringing great help to the extraction of useful information, NLP technology also raises security concerns. In this paper, we present how text adversarial attacks can cause security problems in IIoT based smart grid, which may lead to serious consequences in some scenarios. Specifically, we develop the Important Sentences Perturbed and Encoder/Decoder (ISPED), a novel text adversarial attack algorithm for natural language classification models on the sentence-level. We select sentences that have more influence on the results to disturb while keeping the semantics basically unchanged to reduce smart grid workers’ perception of the attack. Experiments on different datasets and models show that our attacking method can effectively reduce the classification accuracy. Meanwhile, by comparing the original examples with the adversarial examples, we demonstrate that the semantics of the examples remain basically the same.