A tool for textual adversarial attack via multi-objective optimization

Textual adversarial attacks generate adversarial examples that retain similar semantics to the original text and feed them into the target model to detect potential vulnerabilities by comparing output differences. This approach effectively addresses the scarcity of annotated test data during the testing phase. Existing methods often rely on greedy strategies for candidate word selection, which may result in contextually inappropriate or unnatural perturbations, thereby compromising the overall quality of the adversarial examples. To address this issue, we propose MOBTAG, a tool for generating textual adversarial examples based on multi-objective optimization. MOBTAG integrates principles from both multi-objective optimization and genetic algorithms. It improves the attack success rate while maintaining high semantic similarity and readability between adversarial examples and the original texts, thereby enabling the generation of high-quality adversarial examples • A textual adversarial attack tool which generates high-quality adversarial examples via multi-objective optimization. • Evaluation shows it outperforms baselines with higher attack success rates, better textual similarity, and improved readability. • Discussing the limitations of the tool and the potential for further improvement.