Abstract
Natural language processing (NLP) models are known to be vulnerable to adversarial examples, similar to image processing models. Studying adversarial texts is an essential step to improve the robustness of NLP models. However, existing studies mainly focus on analyzing English texts. In this paper, we analyze the differences between Chinese and English and propose Argot, a novel black-box adversarial Chinese text generation solution. Argot utilizes linguistic characteristics such as pinyin, glyphs, and radicals to effectively and efficiently generate adversarial Chinese texts with high readability and success rates.
Introduction
Deep Neural Networks (DNNs) are notoriously fragile when faced with adversarial examples—inputs specifically designed to trigger misclassification. While this phenomenon is well-studied in English NLP, Chinese presents unique challenges. In English, minor perturbations like swapping a letter often leave a word recognizable. In Chinese, characters are "sememes" (the smallest units of meaning). Deleting or changing a single character can completely flip a sentence's meaning. Argot addresses this by creating an attack strategy tailored to the phonetic and visual nuances of the Chinese writing system.
The Argot Framework
Argot operates in a black-box setting, requiring only access to the target model’s output labels and confidence scores. The workflow consists of two main phases:
- Identifying Important Words: The system iteratively removes words from a sentence to identify which are most critical to the classification based on the drop in model confidence.
- Applying Perturbations: Argot applies five types of mutations to these critical words to generate adversarial samples.
Linguistic Perturbation Strategies
To maintain readability while fooling the model, Argot employs several Chinese-specific mutation types:
- Pinyin-based Attacks: These exploit common typos by replacing characters with homophones or those with similar pinyin representations (e.g., nasal shifts). This was found to be the most effective strategy because most models are trained on discrete tokens rather than phonetic features.
- Glyph-based Attacks: This method replaces a character with one that looks visually similar. The authors trained a Siamese Network to identify characters with similar visual structures that humans might easily overlook as typos.
- Splitting-character: Breaking a character into its constituent radicals (e.g., splitting 赶 into 走 and 干). This mimics how a reader might perceive a character even if it is "broken" into its components.
- Synonyms and Shuffling: Standard techniques involving replacing words with similar meanings or swapping character order within multi-character words.
Experimental Results
The authors evaluated Argot on sentiment and news classification datasets. The framework achieved a 97.7% success rate for non-targeted attacks and 98.73% for targeted attacks. Remarkably, these results were achieved by changing less than 12% of the characters.
A user study with 25 native speakers confirmed the high readability of the generated texts: 84% found the adversarial texts readable, and 91% correctly identified the original sentiment or category. This demonstrates a "robustness gap" where humans easily "auto-correct" visual or phonetic errors that cause DNNs to fail.
Conclusion
Argot demonstrates that Chinese NLP models are highly vulnerable to attacks that exploit the language's unique structure. While character-based models showed more robustness than word-based models, both remain susceptible. Future work must focus on incorporating phonetic and visual information into model embeddings to build more resilient Chinese NLP systems.
Summary updated