Low-Query Adversarial Sample Generation Strategy for Tibetan Text Classification

In an effort to resolve the problem of diminished covertness arising from frequent inquiries during black-box adversarial assaults on Tibetan text classification models, this academic paper puts forward a creative means of generating adversarial samples with low query requirements, which is built upon a masked language model. This approach fuses the Tibetan-BERT model with a syllable-level substitution scheme and deploys rapid perturbation and progressive replacement maneuvers to slash the query numbers and augment the success odds of attacks. Through experimentation on two publicly accessible Tibetan text classification datasets, it becomes evident that the proposed methodology eclipses existing ones in both attack success rate and query efficiency. It hits an astounding 98.8% attack success rate while also conspicuously reducing the number of queries. These experimental findings attest that the proposed method doesn't just uphold semantic coherence but also markedly amplifies attack efficiency. Additionally, it's proven that the proposed adversarial attack modality can proficiently seize the semantic idiosyncrasies of Tibetan text and substantially hike query efficiency, permitting triumphant attacks on Tibetan text classification models with a lesser number of queries.