随着人工智能技术的飞速发展，大型语言模型的应用日益广泛，其安全问题也日益受到关注。近日，百度安全副总经理冯景辉在AICon全球人工智能大会上海站上，分享了百度在大型语言模型安全构建方面的实践和创新。
冯景辉指出，大型语言模型的智能性、不确定性和不可解释性为内容安全带来了重大挑战。为此，百度在模型设计阶段就深入考虑安全性问题，并提出了数据清洗、安全对齐、内生安全技术以及安全围栏等措施，形成了一套完整的安全解决方案。
在数据清洗方面，百度采用了四步法，包括数据集评估、隐私脱敏、内容合规清洗、完整性评估，以确保数据的安全性。同时，百度还引入了代答模型，以提高内容审核的自动化和智能化水平。
冯景辉强调，构建原生安全的重要性，通过有监督微调和人类反馈强化学习等技术，可以显著提升模型的安全性和可靠性。
在关键阶段的安全挑战方面，百度采取了具体的安全措施，包括在训练阶段进行数据清洗，在精调阶段进行安全对齐，在推理和部署阶段实现数据安全，以及在业务运营阶段防范模型生成内容的安全性风险。
此外，百度还推出了大模型数据安全解决方案，通过密态数据训练、模型文件加密流转等技术，实现了大模型零信任、零改造的全流程解决方案。
冯景辉的分享为业界提供了大型语言模型安全构建的实践参考，有助于推动人工智能技术的健康发展。
英语如下：
News Title: “Baidu’s Feng Jinghui Reveals: How to Build a Native Security System for Large Models”
Keywords: Baidu, Security Construction, Large Models
News Content:
As artificial intelligence technology advances at a rapid pace, the application of large language models is becoming increasingly widespread, and so too is the growing concern over their security issues. Recently, Feng Jinghui, Vice President of Security at Baidu, shared Baidu’s practical experiences and innovations in the security construction of large language models at the AICon Global Artificial Intelligence Conference in Shanghai.
Feng Jinghui pointed out that the intelligence, unpredictability, and lack of interpretability of large language models pose significant challenges to content security. To address these issues, Baidu has considered security concerns deeply from the model design stage and proposed measures such as data cleaning, safety alignment, intrinsic security technology, and safety barriers, forming a comprehensive security solution.
In terms of data cleaning, Baidu employs a four-step method including dataset evaluation, privacy de-identification, content compliance cleaning, and integrity evaluation to ensure the safety of the data. Additionally, Baidu has introduced answer models to enhance the automation and intelligence of content review.
Feng Jinghui emphasized the importance of building native security, stating that technologies such as supervised fine-tuning and human-in-the-loop reinforcement learning can significantly improve the safety and reliability of the models.
In terms of key stages of security challenges, Baidu has taken specific security measures, including data cleaning during the training phase, safety alignment during the fine-tuning phase, data security during inference and deployment, and the prevention of security risks in the content generated by models during business operations.
Moreover, Baidu has launched a large model data security solution, utilizing technologies such as privacy-preserving data training and encrypted model file flow to achieve a full-process zero-trust, zero-modification solution for large models.
Feng Jinghui’s presentation provides industry reference for the security construction of large language models, helping to promote the healthy development of artificial intelligence technology.
【来源】https://mp.weixin.qq.com/s/04Yr9Xsv1HpdLLgKifC2KA
Views: 2