Research On Attack Threat Of Emojis For Text Classification Tasks

As the number of emojis and the application scenarios of emojis continue to increase,it is inevitable to use Natural Language Processing technology to analyze massive emojis data.However,as a mainstream Natural Language Processing technology,neural network model itself is vulnerable to adversarial attacks and backdoor attacks.As emojis receive more and more attention,it is important to ensure their safety in applications.Therefore,whether emojis can be used as adversarial attack perturbation or backdoor attack trigger to pose attack threats to different neural network classification models becomes an important research topic.In view of the above problems,this paper focuses on the possible attack threats caused by emojis on text classification tasks,and the main work is as follows.Firstly,in order to explore the possible adversarial attack threats caused by emojis to text classification tasks,this paper takes advantage of the characteristic that emojis can express rich semantics,and proposes two adversarial sample generation strategies based on different ways of emojis expressing semantics: the replacement and insertion strategy based on emotional similar emojis SSESI,and the insertion strategy based on text specific semantic emojis TSSEI.Meanwhile,an adversarial attack model based on emojis is constructed.Both methods do not make any modifications to the text content in the samples,but only generate adversarial samples by perturbing emojis.The purpose is to explore the impact of using emojis as perturbation alone on text classification tasks through different emojis semantic expression scenarios.Secondly,in order to explore the possible backdoor attack threats caused by emojis to text classification tasks,this paper takes advantage of the feature that emojis can be inserted into any position in the text,and proposes three trigger construction strategies based on different ways of processing emojis in classification tasks: unknown token trigger construction strategy based on emojis EUT,tokenized trigger construction strategy based on specific emojis EST,and combined trigger construction strategy based on emojis and text ETT.Meanwhile,a backdoor attack model based on emojis is constructed.This purpose is to explore the impact of using emojis to construct triggers on text classification tasks through different processing scenarios.Finally,the effectiveness experiments of different attack strategies are carried out on various datasets,so as to verify the attack threats caused by emojis to text classification tasks and summarize the characteristics of attacks using emojis according to the experimental results.