Help Center/ MaaS/ Security and Compliance/ Model Security
Updated on 2026-09-20 GMT+08:00
View PDF
Share

Model Security

For model inference input and output, content compliance review and risk interception capabilities are provided, covering both self-operated built-in models and third-party models from KooGallery.

Risk Control Capabilities Provided by MaaS

In addition to the security advantages of the computing backbone, MaaS also provides various risk control capabilities to ensure model security.

Table 1 Risk control capabilities provided by MaaS

Capability

Description

Native model security

The built-in security capabilities of the model are provided by the model vendor and take effect by default. No configuration is required.

ModelArts Guard

MaaS provides additional input/output harmful content interception in the inference pipeline, supporting text, image, and video content moderation. Text moderation covers sensitive words and semantic review of politics, sensitive information, terrorism, prohibited items, abuse, and advertisements.

ModelArts Guard switch
CAUTION:

Enabling ModelArts Guard introduces additional latency. Disabling it may fail to block some violations.

Customers at V2 or higher levels can disable ModelArts Guard. After it is disabled, only native model security is applied. The procedure is as follows:

  • Go to the Real-Time Inference built-in services page, and choose More > Disable ModelArts Guard in the Operation column.
  • Alternatively, select ModelArts Guard when creating or editing an endpoint.

Risk identification

When the model detects that the input or output content may contain sensitive information, it returns the HTTP 403 status code and error code ModelArts.81011 to the client for identification and processing.

Risk Control Detection Method

ModelArts Guard detects the input and output phases in the inference pipeline. The process is as follows.

  • Input detection: After a user request reaches the MaaS gateway, the user input is sent to ModelArts Guard for detection. If the request matches a risk rule, the model is not called.
  • Output detection: After the input passes the detection, the model is called normally. The content returned by the model is sent for detection again. If the content matches a rule, the model output is processed accordingly.
  • Handling policy: If the request matches a risk control rule, it is blocked by default (the error code ModelArts.81011 is returned). If the request does not match any risk control rule, it is returned normally.

Risk Control Policies for Preset Models

  • By default, the inference pipeline enables ModelArts Guard, which performs bidirectional detection on both input and output.
  • If risky content is detected, the request is intercepted and error code ModelArts.81011 is returned.
  • Customers at V2 or higher levels can disable ModelArts Guard. After it is disabled, native model security is applied.

Security Capabilities

MaaS has built a hierarchical security protection system for the inference pipeline, categorized into the following three types of capabilities:

  • Network Security Layer

    For public network call chains, the MaaS gateway provides intrusion prevention and traffic cleaning capabilities to intercept malicious requests and attack traffic:

    • WAF: protects against injection, crawlers, and malicious payloads, and focuses on large packets and ultra-long context features.
    • AAD and Anti-DDoS: ELB AAD protects against DDoS attacks.
    • APIG: provides a unified entry and centralized TLS, rate limiting, concurrency control, access control, and monitoring.
  • Security Detection and Response Layer

    Provides threat detection, vulnerability management, and event source tracing capabilities to ensure the overall security posture of the platform:

    • Vulnerability scanning and situational awareness: proactively identifies security threats and intrusions and generates alarms.
    • SecMaster: aggregates security logs, displays threat situations, and supports end-to-end backtracking and locating of security events.
  • Security Audit and Monitoring Layer

    Provides audit, metric monitoring, and exception alarm capabilities for operation behaviors and runtime metrics. For details, see Operation Audit and Model Monitoring.

Risk Control Disclaimer

You should be aware of the following information about content moderation and risk control:

  • The native security of models is provided by the model vendors. The native security capabilities of models (including open-source models and third-party models) on the MaaS platform are provided by the model vendors. Huawei Cloud does not guarantee the compliance, accuracy, or integrity of the native security of models. For open-source models, Huawei Cloud does not assume any responsibility for their native security capabilities.
  • ModelArts Guard cannot ensure 100% filtering and compliance. ModelArts Guard provided by MaaS is an auxiliary risk identification and blocking capability based on existing technical means. Due to the objective limitations of content risk control technologies, it cannot ensure that all violations can be identified and blocked. You should independently review the compliance of the input and output content based on your service scenarios. Do not use the ModelArts Guard as the only compliance assurance method.
  • Third-party model disclaimer: Third-party models and their outputs do not represent Huawei Cloud's opinions. Huawei Cloud does not guarantee the compliance, security, authenticity, accuracy, or integrity of third-party models. Any damage caused by third-party models shall be resolved between you and the third-party vendor.
  • Risks of disabling ModelArts Guard: After ModelArts Guard is disabled, you will be liable for any legal responsibilities and third-party claims arising from the model outputting prohibited content. It is recommended that you disable ModelArts Guard with caution in the production environment and build your own content moderation mechanism.

For details, see MaaS Service Agreement.

FAQs

Q1: What are the impacts of disabling ModelArts Guard?

After ModelArts Guard is disabled, only the native security capabilities of the model will be used, which may fail to intercept some prohibited content. You will be liable for any legal responsibilities arising from the model outputting prohibited content.

Q2: What should I do if the content is blocked by ModelArts Guard?

When the content is blocked, the model returns the HTTP status code 403 and error code ModelArts.81011. You can modify the request content (such as the prompt) and try again.