An Effective Adversarial Text Attack through a Block-Sparse Approach with Hamiltonian Insights

Recent research on adversarial attacks shows that natural language processing models are vulnerable to adversarial examples. Current methods usually prove to be inadequate to successfully mislead the model while maintaining high semantic retention and fluency of the adversarial examples. These methods have difficulty in finding the global optimal solution among various, feasible possibilities of minor changes. Consequently, developing textual adversarial attacks that can produce high-quality adversarial examples continues to be a challenging endeavor. This paper proposes Hamiltonian text adversarial attack (HTAA); primarily, it generates a certain number of adversarial candidates with the guidance of limiting perturbations and gradient projections, expanding the searching scope. Besides, it utilises Hamiltonian simulation to efficiently select the final adversarial example that is closer to the global optimum among the candidates. Extensive experiments show that the proposed method improves semantic similarity and fluency, raising by 5% and 12.3% an average respectively; at the same time, it maintains a high successfully attack rate of more than 90% on four different datasets against the target model.