The Large Language Model Security Certification is conducted based on the WDTA AI STR-02 standard. It aims to ensure that large language models can withstand various adversarial attacks, reduce potential risks, and promote their responsible application in real-world scenarios by establishing strict security testing standards and evaluation procedures. The certification includes comprehensive evaluations of models facing various attack types during pre-training, fine-tuning, and inference stages (such as data poisoning, backdoor attacks, instruction hijacking, and prompt obfuscation).

By setting clear evaluation metrics such as attack success rate (R) and refusal rate (D), the certification provides organizations with a scientific and systematic security detection method, helping them identify and fix potential vulnerabilities in the models, thereby enhancing the overall security and reliability of the models.

Additionally, the evaluation emphasizes protective measures for large language models in terms of data privacy, model integrity, and contextual applicability, ensuring that they can produce safe and reliable results when faced with adversarial examples.