GB/T 45958-2025 in English

Standard Overview and Development Background

GB/T 45958-2025, "Cybersecurity Technology - Security Framework for Artificial Intelligence Computing Platforms," is my country's first national standard specifically addressing the security of artificial intelligence computing platforms. Proposed and coordinated by the National Cybersecurity Standardization Technical Committee (SAC/TC260), it was jointly drafted by over 60 organizations, including Huawei Technologies Co., Ltd. and the China Electronics Standardization Institute. Published in 2025, the standard aims to address new security challenges arising from the widespread application of artificial intelligence technology across various fields.

With the rapid development of artificial intelligence technology and the large-scale deployment of specialized hardware devices such as artificial intelligence servers and artificial intelligence accelerators, traditional cybersecurity frameworks are no longer able to fully address the unique security requirements of AI computing platforms. This standard was developed against this backdrop to provide comprehensive security guidance for the design, construction, application, and operation and maintenance of AI computing platforms.

Analysis of the Core Security Framework

The standard establishes a three-dimensional security framework for artificial intelligence computing platforms, including three core components: security functions, security management, and role security responsibilities.

Detailed explanation of resource-layer security functions

Resource-layer security is the foundation of AI computing platform security, covering security requirements in four areas: computing resources, storage resources, network resources, and virtual resources.

Computing Resource Security

In view of the particularity of AI acceleration processors and AI servers, the standard proposes 7 specific security requirements:

• Firmware integrity and authenticity verification mechanism
• Version anti-rollback protection
• Integrity verification interface provision
• Management interface security protection
• Fault monitoring and disposal
• Resource security isolation
• Trusted execution environment construction

Storage Resource Security

In view of the large amount of checkpoint data and model parameters generated during model training and inference, the standard requires:

• Redundant storage to prevent data loss
• Distributed architecture to improve availability
• Multiple backup and anti-ransomware encryption
• Cryptographic technology to protect confidentiality and integrity

Implementation of Scheduling Layer Security Functions

Scheduling layer security ensures the reliability and security of resource allocation and task execution, including resource scheduling security and task scheduling security.

Application Support Layer Security Practices

Application support layer security focuses on the security of three key links: data processing, model training, and model reasoning.

Data Processing Security

The standard innovatively proposes the concept of a dataset bill of materials, requiring:

• Generate a dataset bill of materials file and transfer it with the dataset
• Support integrity, authenticity, and version correctness verification
• Implement security defect tracing
• Protect the confidentiality and integrity of important data

Model Training Security

For the machine learning model training process, the following requirements apply:

• Generate a model bill of materials file
• Retain tamper-proof log records
• Support compliance auditing and problem tracing

Model Reasoning Security

For the model reasoning link, four core security measures are proposed:

• Model authorization control and encryption and decryption protection
• Verification of integrity, authenticity, and version correctness
• Security testing of reasoning requests
• Security testing of reasoning results

Construction of a security management system

The standard constructs a complete security management system from seven dimensions, which complements the security functions.

Division of security responsibilities

The standard clarifies the security responsibilities of four types of participants and builds a collaborative protection system.

Responsibilities of the Platform Provider

As the infrastructure provider, assume the most comprehensive security responsibility:

• Provide technical support for security functions
• Implement a security management system
• Ensure compliance with password application
• Implement personal information protection

Responsibilities of the Data Provider

Focus on the security of the data supply link:

• Provide secure data sets
• Support verification and traceability
• Support transparent auditing

Responsibilities of the Model Provider

Responsible for model development and supply security:

• Comply with data security requirements
• Provide security models or interfaces
• Support model verification and auditing

Responsibilities of the Application Provider

Assume ultimate application security responsibility:

• Comply with general security requirements
• Protect model and data security
• Interface security protection
• Attack identification and blocking
• Logging and tracing

Implementation recommendations and best practices

Based on the standard requirements, the following implementation recommendations are proposed:

Technical implementation level

• Establish a firmware security management system for AI accelerated computing resources to implement integrity verification and version control
• Deploy a distributed storage architecture to ensure checkpoint data redundancy and availability
• Implement network partitioning and isolation to distinguish between data processing, model training, and inference networks
• Adopt trusted execution environment technology to protect in-memory model and data security

Management implementation level

• Develop a bill of materials management system to achieve full lifecycle traceability of data and models
• Establish a multi-role collaborative security mechanism to clarify responsibility boundaries and collaboration processes
• Implement hierarchical log management to meet audit requirements of different granularities
• Carry out regular security assessments to promptly discover and fix vulnerabilities

Compliance implementation level

• Follow the requirements of the Level Protection System and make adaptive adjustments based on AI characteristics
• Implement personal information protection regulations to ensure legal and compliant data processing
• Adopt national standard cryptographic algorithms to ensure encryption security strength
• Establish a compliance audit mechanism and conduct regular security compliance inspections

Standard evolution and future prospects

GB/T 45958-2025, as the first national standard in the field of artificial intelligence computing platform security, reflects my country's forward-looking thinking in AI security governance. With the rapid development of AI technology, future standards may need to:

• Adapt to the security requirements of new AI architectures such as large models
• Cover emerging technologies such as federated learning and privacy-preserving computing
• Strengthen coordination and mutual recognition with international standards
• Continuously update security requirements to respond to new threats

The implementation of this standard will effectively promote the healthy and orderly development of my country's artificial intelligence industry and lay a solid foundation for building a secure and trustworthy AI ecosystem.