TL;DR
This paper introduces TSAttacker, the first syllable-level black-box textual adversarial attack specifically designed for Tibetan language models. The research highlights the vulnerability of these models to such attacks and underscores the need for increased focus on the robustness of Chinese minority language processing.
What Problem Does This Paper Address?
The field of Natural Language Processing (NLP) has seen significant advancements, but neural network models remain susceptible to adversarial attacks. These attacks involve making subtle, imperceptible changes to input text that can cause NLP models to misclassify or misinterpret the content. While research on textual adversarial attacks has been extensive for English and to some extent for Chinese, there's a notable gap in research concerning Chinese minority languages. This lack of attention poses a significant risk to the development and security of information processing systems for these languages, potentially impacting everything from user experience to data integrity. The paper specifically targets Tibetan, a language with unique linguistic characteristics, to address this research gap.
Methods
The core contribution is the TSAttacker algorithm, a syllable-level black-box textual adversarial attack. Unlike character or word-level attacks, TSAttacker operates at the syllable level, which is a natural granularity for Tibetan. The attack proceeds in two main stages:
- Syllable Substitution: The algorithm identifies candidate syllables for substitution using syllable cosine distance. It leverages pre-trained Tibetan syllable embeddings and selects syllables whose embeddings are within a specified cosine distance (
dmax) from the original syllable's embedding. This allows for controlled similarity between original and substituted syllables. - Substitution Order Scoring: To determine the optimal order of syllable substitutions, TSAttacker employs a scoring mechanism. This mechanism combines syllable saliency (how much changing a syllable affects the model's classification probability) with a probability-weighted word saliency approach. The goal is to find the substitution that causes the most significant change in the model's output with the highest attack effect.
The attack is "black-box," meaning it doesn't require access to the model's internal architecture or gradients, making it more practical for real-world scenarios.
Datasets
The paper utilizes the following datasets for evaluation:
- TNCC-title: A news title classification dataset for Tibetan, with 9,276 samples and an average of 16 syllables per title.
- TNCC-document: A news document classification dataset for Tibetan, with 9,204 samples and an average of 689 syllables per document.
- TU_SA: A Tibetan sentiment classification dataset, consisting of 10,000 samples with an average of 28 syllables per text.
These datasets are used to fine-tune pre-trained language models (PLMs) for downstream tasks.
Key Insights and Ideas
- Syllable-level attacks are effective for Tibetan: The paper demonstrates that targeting syllables, a natural linguistic unit in Tibetan, is a viable and effective strategy for adversarial attacks.
- Chinese minority language models are vulnerable: The research clearly shows that even state-of-the-art Chinese minority PLMs (specifically, CINO) are susceptible to textual adversarial attacks, indicating a significant robustness gap.
- Language-specific attacks are crucial: The unique linguistic structure of Tibetan necessitates specialized attack methods, as generic approaches might not be as effective.
- Model robustness is not solely dependent on size: The experiments suggest that larger models don't always exhibit better robustness, indicating that architectural choices and training data play a critical role.
- Standard language formatting improves performance: Fine-tuning models on texts that adhere to standard Tibetan script conventions (using 'tsheg' between syllables) leads to better performance compared to models trained on texts with spaces or no separators.
Key Results
TSAttacker was tested on six fine-tuned CINO models across the TNCC-title, TNCC-document, and TU_SA datasets. The key findings include:
- High Attack Success Rate (ASR): TSAttacker achieved high ASRs, ranging from 63.80% to 76.05% across different models and datasets.
- Significant Accuracy Drop Value (ADV): The attack caused substantial drops in model accuracy, with ADV values ranging from 0.2240 to 0.3859.
- High-Quality Adversarial Samples: The generated adversarial samples had a low average Levenshtein Distance (LD), indicating minimal edits and thus higher quality. For TNCC-title, LD was around 1.6-1.7; for TU_SA, it was around 2.7-2.9. For TNCC-document, LD was significantly higher (around 39-41), likely due to the longer text length.
- Model Robustness Improvement Needed: The results consistently show that the tested models have considerable room for improvement in terms of robustness.
Limitations and Future Work
- Limited Scope: The paper acknowledges that its work "only preliminarily explores the field" and focuses on the Tibetan part of the first Chinese minority multilingual PLM.
- Future Research Directions: The authors suggest that future work should:
- Develop attack methods that better incorporate the linguistic characteristics of Chinese minority languages.
- Focus on improving the overall robustness of Chinese minority PLMs, addressing the significant gap in corpus size compared to English and Chinese.
- Explore posterior defense mechanisms against the proposed textual adversarial attacks.
- Continue to concentrate on the security issues faced in Tibetan information processing.
Summary updated