Adversarial Attacks Based on Sentiment Analysis Under the BERT Model

With the wide application of deep learning technology in sentiment analysis tasks, its adversarial vulnerability has gradually become a research hotspot. This study focuses on the adversarial attack methods for deep learning sentiment analysis models and systematically analyses the white-box attacks and their practical effects. We experimentally validate the efficacy of FGSM attacks using the IMDB movie review corpus. The results demonstrate that the deep learning sentiment analysis model exhibits significant vulnerability in the face of adversarial attacks. When the perturbation size (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\varepsilon$</tex>) is increased from 0 to 0.1, the model accuracy dramatically decreases from 92.5% to 30.0%, while the attack success rate significantly increases from 0% to 70.0%. In addition, the experiments reveal the trade-off between perturbation size and steganography for the attack methods. As the value of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\varepsilon$</tex> increases, the semantic similarity (BERTScore) decreases from 1.0 to 0.75, indicating that the attack effectiveness improves while the covertness decreases. The analysis in this paper can provide theoretical support for subsequent defence studies and quantitative data reference for model security in practical applications. Future research can further explore more efficient attack methods and propose more effective defence strategies on this basis.