2026/07/06 by Yuming Lu, Jie Zhang, Junhua Zhang +1
Computer Science · #Adversarial Robustness in Machine Learning #Topic Modeling #Advanced Graph Neural Networks
paper · doi:10.1093/comjnl/bxag074
Abstract Adversarial attacks involve introducing small perturbations to the original text to mislead models. Existing adversarial attack methods often rely on fixed search or substitution strategies, which may ignore the cumulative effects of previous attack steps and make limited use of attack history. This paper presents CBSL, a search-and-learning framework for Chinese word-level black-box adversarial sample generation. First, we design a simulated annealing search framework that incorporates a multi-objective evaluation function and an adaptive, threshold-controlled iterative strategy for adversarial attacks. Second, to fully account for the contextual semantics of the attack position, CBSL employs a pretrained masked language model to generate replacement words and utilizes a dynamic replacement mechanism to produce candidate adversarial samples. Finally, we fine-tune this generative model using attack history to reduce the uncertainty caused by candidate generation and search order, thereby improving both attack efficiency and adversarial sample quality. A composite loss function combining masked-position prediction and semantic preservation is designed to guide this fine-tuning process. Experimental results show CBSL can effectively produce high-quality Chinese adversarial samples. On both THUCNews and CAIL2018 datasets, CBSL achieves higher attack success rates than representative baseline methods, with improvements of ~10%–20% in most settings, while maintaining high semantic similarity. Additionally, we conduct ablation studies, adversarial training experiments, transferability analysis, and attack preference analysis. The results further confirm the efficiency and effectiveness of our method in generating high-quality adversarial samples.