9月8日，全球公共安全合作论坛（连云港）2023年大会数据安全分论坛在北京召开，中国电子党组成员、副总经理，中国电子数据产业有限公司董事长、50人论坛委员陆志鹏出席会议并作主旨发言。
陆志鹏指出，数据安全是基于网络安全和信息安全的新时代国家安全的关键支点和核心内涵，在人工智能逐步落地应用，大模型发展势头强劲，数据安全问题日益凸显的当下，着力解决数据安全突出问题，有效提升数据安全治理能力显得至关重要。
数据是支撑大模型发展的重要引擎，传统大模型训练的数据集往往存在数据未经过治理组织、数据使用缺少合规审查和安全保护等不足，造成用户隐私和企业商业数据泄露、违反法律法规相关要求、侵权等数据安全、数据合规安全及产权保护问题。
中国电子联合清华大学及相关机构，开展基础科学研究及理论实践，着力构筑大模型发展的安全可信数据底座，首创数据元件理念，以数据元件“中间态”将原始数据和数据应用进行隔离，依靠对数据元件的安全审查，实现精准监管，有效破除大模型数据安全风险；并探索形成了基于数据元件的新型技术架构、文本元件加工流程和基于数据元件的大模型微调及训练验证模式，实现对原始数据的有效保护、合规处理及大模型的高效、精准训练。
翻译：
Lu Zhipeng: Thinking and exploration of data security under the big model boom
On September 8, the Data security sub-forum of the Global Public Security Cooperation Forum (Lianyungang) 2023 Conference was held in Beijing, and Lu Zhipeng, a member of the Party Group of China Electronics, deputy general manager, chairman of China Electronic Data Industry Co., LTD., and a member of the 50-member forum, attended the meeting and made a keynote speech.
Lu Zhipeng pointed out that data security is the key fulcrum and core connotation of national security in the new era of network security and information security, in the gradual application of artificial intelligence, the development of large models is strong, and data security problems are becoming increasingly prominent at present, it is crucial to solve prominent problems of data security and effectively improve data security governance ability.
Data is an important engine to support the development of large models. Data sets trained by traditional large models often have deficiencies such as data not being governed by the organization, lack of compliance review and security protection in data use, resulting in user privacy and business data leakage, violation of relevant requirements of laws and regulations, infringement and other data security, data compliance security and property rights protection.
China Electronics, together with Tsinghua University and relevant institutions, carries out basic scientific research and theoretical practice, focuses on building a safe and trusted data base for the development of large models, and creates the concept of data components, isolating the original data and data applications with the “intermediate state” of data components, and realizing accurate supervision and effectively eliminating the security risks of large model data by relying on the security review of data components. And explored the formation of a new technical architecture based on data elements, text elements processing flow and large model fine-tuning and training verification mode based on data elements, to achieve effective protection of original data, compliance processing and efficient and accurate training of large models.
