随着人工智能领域大模型(large model)的广泛应用,大模型,尤其是大语言模型(large language model,LLM)的安全问题受到了广泛关注.大模型作为一种新兴技术,与之相关的安全态势分析以及安全体系建设均亟待挖掘与探索.从社会关系以及技术...随着人工智能领域大模型(large model)的广泛应用,大模型,尤其是大语言模型(large language model,LLM)的安全问题受到了广泛关注.大模型作为一种新兴技术,与之相关的安全态势分析以及安全体系建设均亟待挖掘与探索.从社会关系以及技术应用2个视角,分析了大模型安全的整体趋势.同时,基于大模型自身的特点,梳理了大模型安全能力建设的实践思路,为大模型研发、大模型应用构建提供了安全体系构建的参考方案.介绍的大模型安全能力实践方案包括安全评估基准建设、模型价值观对齐方法、模型线上服务安全系统建设3个部分.展开更多
大模型红队测试(Large Model Red Teaming)旨在让大语言模型(Large Language Model, LLM)接收对抗测试,从而诱使模型输出有害的测试用例,进而发现模型中的漏洞并提高其鲁棒性。大模型红队测试是大模型领域的前沿课题,近年来受到学术界...大模型红队测试(Large Model Red Teaming)旨在让大语言模型(Large Language Model, LLM)接收对抗测试,从而诱使模型输出有害的测试用例,进而发现模型中的漏洞并提高其鲁棒性。大模型红队测试是大模型领域的前沿课题,近年来受到学术界和工业界的广泛关注。研究者们针对大模型红队测试提出了众多解决方案,并在模型对齐上取得了一定进展。然而,受限于大模型红队数据的短缺和评价标准的模糊,现有研究大多局限于针对特定的场景进行评估。文中首先从与大模型安全相关的定义出发,对其所涉及的各种风险进行阐述;其次,针对大模型红队测试的重要性及其主要类别进行了阐述,综述和分析了相关红队技术的发展历程,并介绍了已有的数据集和评价指标;最后,对大模型红队测试的未来发展趋势进行了展望和总结。展开更多
The focus of this paper is the ill-conditioned problems in the dam safety monitoring model. The reasons to give rise to the ill-conditioned problems in statistical models,deterministic models and hybrid models are ana...The focus of this paper is the ill-conditioned problems in the dam safety monitoring model. The reasons to give rise to the ill-conditioned problems in statistical models,deterministic models and hybrid models are analyzed in detail,and the criterions for ill-conditioned models are investigated. It is shown that safety monitoring models are not easy to be ill-conditioned if the number of influence factors is less than seven. Moreover,the models have a high accuracy and can meet the engineering requirements. Another frequently encountered problem in establishing a safety monitoring model is the existence of inflection points,which are often present in the mathematical model for the hydraulic components in deterministic models and hybrid models. The conditions for inflection points are studied and their treatments are suggested. Numerical example indicates that the treatments proposed in this paper are effective in removing the ill-conditioned problems.展开更多
This paper presents a model to simulate the safe behavior of Dagangshan arch dam with a rate-dependency anisotropic damage model. This model considers the damage of asymmetry and anisotropy under cyclic loading of ten...This paper presents a model to simulate the safe behavior of Dagangshan arch dam with a rate-dependency anisotropic damage model. This model considers the damage of asymmetry and anisotropy under cyclic loading of tension and compression, and it is used in the compiled finite element code. The material parameters used in the model can be identified from uniaxial static and dynamic experiments. Thereafter, it is used for analyzing damage and failure patterns of the dam subjected to water pressure and strong earthquakes. The numerical results show that it is necessary to consider both asymmetry between tension and compression and anisotropy of damage. Severe damage regions of the dam reveal brittle and risky positions clearly. Meanwhile damage patterns show the failure trend and safety behaviors of the dam. These results match well with that of the experiments carried out in DUT. The proposed model may be used to predict the damage patterns and potential failure modes of concrete structures like the dam. And the aseismic performance of the dam can be figured out.展开更多
文摘随着人工智能领域大模型(large model)的广泛应用,大模型,尤其是大语言模型(large language model,LLM)的安全问题受到了广泛关注.大模型作为一种新兴技术,与之相关的安全态势分析以及安全体系建设均亟待挖掘与探索.从社会关系以及技术应用2个视角,分析了大模型安全的整体趋势.同时,基于大模型自身的特点,梳理了大模型安全能力建设的实践思路,为大模型研发、大模型应用构建提供了安全体系构建的参考方案.介绍的大模型安全能力实践方案包括安全评估基准建设、模型价值观对齐方法、模型线上服务安全系统建设3个部分.
文摘大模型红队测试(Large Model Red Teaming)旨在让大语言模型(Large Language Model, LLM)接收对抗测试,从而诱使模型输出有害的测试用例,进而发现模型中的漏洞并提高其鲁棒性。大模型红队测试是大模型领域的前沿课题,近年来受到学术界和工业界的广泛关注。研究者们针对大模型红队测试提出了众多解决方案,并在模型对齐上取得了一定进展。然而,受限于大模型红队数据的短缺和评价标准的模糊,现有研究大多局限于针对特定的场景进行评估。文中首先从与大模型安全相关的定义出发,对其所涉及的各种风险进行阐述;其次,针对大模型红队测试的重要性及其主要类别进行了阐述,综述和分析了相关红队技术的发展历程,并介绍了已有的数据集和评价指标;最后,对大模型红队测试的未来发展趋势进行了展望和总结。
基金supported by the National Natural Science Foundation of China (Grant Nos. 51079046, 50909041, 50809025, 50879024, 51139001)the National Science and Technology Support Plan (Grant Nos. 2008BAB29B03, 2008BAB29B06)+5 种基金the Special Fund of State Key Laboratory of China (Grant Nos. 2009586012, 2009586912, 2010585212)the Fundamental Research Funds for the Central Universities (Grant Nos. 2009B08514, 2010B20414, 2010B01414, 2010B14114)China Hydropower Engineering Consulting Group Co. Science and Technology Support Project (Grant No. CHC-KJ-2007-02)Jiangsu Province "333 High-Level Personnel Training Project" (Grant No. 2017-B08037)the Graduate Innovation Program of Universities in Jiangsu Province (Grant No. CX09B_163Z)the Science Foundation for the Excellent Youth Scholars of Ministry of Education of China (Grant No. 20070294023)
文摘The focus of this paper is the ill-conditioned problems in the dam safety monitoring model. The reasons to give rise to the ill-conditioned problems in statistical models,deterministic models and hybrid models are analyzed in detail,and the criterions for ill-conditioned models are investigated. It is shown that safety monitoring models are not easy to be ill-conditioned if the number of influence factors is less than seven. Moreover,the models have a high accuracy and can meet the engineering requirements. Another frequently encountered problem in establishing a safety monitoring model is the existence of inflection points,which are often present in the mathematical model for the hydraulic components in deterministic models and hybrid models. The conditions for inflection points are studied and their treatments are suggested. Numerical example indicates that the treatments proposed in this paper are effective in removing the ill-conditioned problems.
基金supported by the National Natural Science Foundation of China (Grant Nos. 90510017 and 50878123)the National Basic Research Program of China (Grant No. 2007CB714104 )+1 种基金the Innovative Project for Postdoctor of Shandong Province (Grant No. 200803037)the Research Project of SUST Spring Bud (Grant No. 2008AZZ107)
文摘This paper presents a model to simulate the safe behavior of Dagangshan arch dam with a rate-dependency anisotropic damage model. This model considers the damage of asymmetry and anisotropy under cyclic loading of tension and compression, and it is used in the compiled finite element code. The material parameters used in the model can be identified from uniaxial static and dynamic experiments. Thereafter, it is used for analyzing damage and failure patterns of the dam subjected to water pressure and strong earthquakes. The numerical results show that it is necessary to consider both asymmetry between tension and compression and anisotropy of damage. Severe damage regions of the dam reveal brittle and risky positions clearly. Meanwhile damage patterns show the failure trend and safety behaviors of the dam. These results match well with that of the experiments carried out in DUT. The proposed model may be used to predict the damage patterns and potential failure modes of concrete structures like the dam. And the aseismic performance of the dam can be figured out.
