On October 11, the official website of the National Information Security Standardization Technical Committee released the "Basic Requirements for Generative Artificial Intelligence Service Security" (Draft for Comments), soliciting public opinions.
Original announcement: https://www.tc260.org.cn/front/postDetail.html?id=20231011143225
This is the country's first regulatory draft specifically for the field of generative AI security . It also supports the "Interim Measures for the Management of Generative Artificial Intelligence Services" launched by seven departments including the Cyberspace Administration of China in July .
The draft for comments proposes for the first time the basic security requirements that generative AI service providers need to follow , covering aspects such as corpus security, model security, security measures, and security assessment .
Some highlights of the document are described below.
Requirement: Providers should not conduct research and development using base models that have not been registered with the competent authority. During the training process, the security of the generated content should be regarded as one of the main considerations for evaluating the quality of the generated results.
In each conversation, the security of user input information should be checked to guide the model to generate positive content .
The "Requirements" proposed that a blacklist of corpus sources should be established, and data from the blacklist sources should not be used for training . Security assessments should be conducted on corpus from each source. If the content of a single source corpus contains more than 5% of illegal and harmful information, it should be added to the blacklist .
When using corpus containing personal information, the authorization and consent of the corresponding personal information subject must be obtained, or other conditions for the legal use of the personal information must be met.
When using corpus that contains biometric information such as faces, the written authorization and consent of the corresponding personal information subject must be obtained, or other conditions for the legal use of the biometric information must be met.
During the training process, the security of the generated content should be regarded as one of the main considerations for evaluating the quality of the generated results.
Full text: https://www.tc260.org.cn/upload/2023-10-11/1697008495851003865.pdf