Full text
- This is an unofficial copy. The document has been archived and reformatted in plaintext for AGORA. Footnotes, tables, and similar material may be omitted. For the official text, visit the original source.
-
This document has been translated from Chinese to English by the Center for Security and Emerging Technology. The translation is unofficial. Translators’ notes and other metadata may be omitted from the AGORA copy. For the original translation, visit https://cset.georgetown.edu/publication/artificial-intelligence-security-standardization-white-paper-2019-edition/.
- This text may be out of date. According to the latest data in AGORA, this document has been proposed, but is not yet enacted or otherwise finalized. This text was collected 2024-10-24 and may have been revised in the meantime. Visit the official source for authoritative text.
5 Recommendations for AI security standardization work
(1) Attach importance to improving a system of AI security standards
We recommend launching AI standardization work, coordinating the planning of a system of AI
security standards, strengthening research into fundamental AI security standards, and deepening the work
of AI application security standards.
The first is the overall planning of a system of AI security
standards. In order to ensure the orderly progress of the development of AI security standards, we
recommend investigating and analyzing the need for AI security standardization in China and giving
priority to research on a system of AI security standards. The system of standards should cover the security
needs of multiple objects such as the foundation, platform, technology, products, and applications of AI
and should be able to clearly define relationships with related standards such as big data security, personal
information protection, cloud computing security, and IoT security.
Second, pay close attention to
research and to implement the AI ethical principles. Focus on the outstanding issues of AI algorithm
discrimination and algorithmic bias. Analyze and review the ethical needs of AI in various scenarios,
develop and refine AI ethical principles, and guide the implementation of principles and requirements
related to AI standards.
(2) Accelerate the development of standards in key areas
AI has the characteristics of wide coverage, complex application scenarios, and many types of security
requirements. We recommend that AI security standardization promotion plans be established to develop
AI security standards in accordance with the concept of “emergency use first, driven by safety incidents.”
The development of standards should be accelerated in key areas, and AI security standardization should
be advanced in an orderly fashion.
First is fundamental AI security standards research. China’s AI security
standards mainly focus on the field of application security, lacking security standards for AI itself or basic
commonality. We recommend that fundamental AI security standards be strengthened with requirements
for AI security aspects such as monitoring and early warning, risk assessment, security accountability, and
research personnel security guidelines based on the New Generation Artificial Intelligence Development
Plan. Standards research should be conducted for such aspects as AI security reference architecture, security
risks, ethical design, and security assessment to grasp AI algorithm security threats and protection needs to
clarify general algorithm security principles and requirements and to strengthen AI algorithm model
security and robustness. AI algorithm models and smart product security requirements and evaluation
methods should be standardized to address the data quality and dataset security issues faced by AI.
Second is a deepening of AI application security standards. Given that AI is becoming ever more integrated with
various application fields, we should begin to study AI products and application security standards. We
recommend that smart product security standards be ahead of AI application standards. AI security
standardization characteristics and requirements should be extracted based on the development of standards
for smart door locks, smart homes, and other fields that have been carried out by the National Information
Security Standardization Technical Committee.
Afterwards, we should give priority to areas where there is
an urgent need for standardization, more mature applications, urgent security needs, wide applications, or
areas that are relatively sensitive. We should develop AI products and application security standards to
improve the security requirements of existing AI standards.
Third is to develop standards according to
the concept of “fulsome research, with priority to urgent needs, driven by security incidents.” We
recommend that priority be given to security standards with pressing security needs and mature applications.
For instance, standards such as Artificial Intelligence Security Reference Framework, Artificial Intelligence
Dataset Security, Artificial Intelligence Data Labeling Security, Machine Learning Algorithm Model
Trustworthiness, Artificial Intelligence Open Source Framework Security, Artificial Intelligence
Application Security Guide, and Artificial Intelligence Security Service Capabilities Requirements should
be prioritized.
Simultaneously, research work on fundamental AI standards should be conducted with the
study and application of security risk assessment standards and key AI product and service security
standards such as smart manufacturing and intelligent networked vehicles. We should gradually promote research on AI security standards in other fields.
(3) Diligently promote the application of AI security standards
In order to improve the effectiveness and operability of AI security standards, address prominent AI
security risks, and explore a path to standardization of the most difficult and pressing issues of AI security,
we recommend that in-depth application and practical work be carried out for AI security standards.
First, improve the pilot mechanisms for AI security standards. Select a number of pilot enterprises to carry
out an evaluation of the applicability of the standards and the effectiveness of implementation, establish the
concept of work of “tracking practice, discovering issues, summarizing experiences, improving standards,
and providing feedback for the next step of standardization,” and promote the rapid and high-quality
development of AI security standardization.
Second, improve the research, promotion, and application
of promotion mechanisms for AI security standards. Organize universities, scientific research institutes,
and enterprises in jointly breaking past AI security standardization difficulties, give play to the advantages
of various institutions, establish AI security standards research, promotion, and application mechanisms
that integrate "industry, university, and research," and promote the benign development of the AI industry.
(4) Effectively strengthen the training of AI security standardization talent
Talent is the cornerstone of AI security standardization work, and we recommend that a multi-level
and multi-type AI security talent training mechanism be established and improved upon.
First, train AI
security professionals. Establish training programs for professional technology, standard setting, publicity
and habituation training (宣贯培训), and testing and evaluation.
Second, encourage universities, research
institutes, and enterprises to establish collaboration. Explore training paths that integrate talent into AI
security.
Third, strengthen support for artificial intelligence security and standardization projects.
Optimize the ratio of scientific research resources and management and evaluation mechanisms to ensure
that relevant talent fields concentrate on solving key AI security issues.
(5) Actively participate in international AI security standardization
International organizations such as ISO and IEEE have organized and undertaken a number of AI
security standardization research tasks and have achieved certain standardization results in some areas. We
recommend that the achievements of international and overseas standardization work be fully digested,
absorbed, and incorporated with China's AI security needs in order to explore AI security standardization
work paths with Chinese characteristics.
First, closely track and study the dynamics and development
trends of AI security standardization work at home and abroad. Consolidate the research achievements
of international AI security standardization, absorb the experiences of standards development overseas, and
promote the better development of AI security standardization in China.
Second, continuously enhance
the influence of China's international standards in the field of AI security. Diligently support the
participation of Chinese institutions and experts in international standardization work, strengthen research
into AI security standards, and encourage Chinese experts to serve in international standardization
organizations and act as editors of international standard projects.
Third, give full play to China's
international standardization exchange and cooperation mechanisms. Given the rich application
scenarios of China's AI industry, standards cooperation and exchange mechanisms in the field of AI security
should be established to enrich the achievements of Chinese AI security standardization work with the help
of international and overseas efforts.
(6) Establish an AI high security risk early warning mechanism as soon as possible
In view of AI technologies, products, and applications with high security risks, we recommend that an
early warning mechanism be studied and proposed for highly dangerous AI security risks.
First, establish
a catalog of highly dangerous AI security risks. Review AI technologies, products, and applications with
outstanding security risks that may create high-impact security issues and classify and categorize risk items
within the catalog.
Second, establish an AI high security risk early warning mechanism. Incorporate
technology, application, and product characteristics and propose an early warning plan for risks.
Third, study and formulate highly dangerous AI safety risk management standards. Starting with standards,
propose a security risk management plan for highly dangerous AI that incorporates techniques, management,
evaluation, and other means and covers such aspects as risk identification, analysis, and processing.
(7) Effectively improve AI security supervision support capabilities
Standardization can strongly support the implementation of AI security supervision, and we
recommend the development of an index evaluation system for AI security standards. First, establish a
sound AI supervision system and formulate supporting standards. The government should use
standards as a powerful starting point to establish a monitoring system that runs through the entire cycle of
AI development, design, data acquisition, and market application to prevent AI from being used illegally
or used in areas that deviate from its intended purpose.
Second, we recommend accelerating research
into AI supply chain security management mechanisms. Develop supporting standards for AI supply
chain security, propose AI supply chain procurement requirements for telecommunications, energy,
transportation, power, finance, and other industries, and promote the piloting of relevant standards in key
areas to provide a useful reference for Chinese party and government departments, and for key industries,
to manage AI supply chain security risks and offer practical guidance for enterprises as they strengthen AI
supply chain management.