Skip to main content
Version: Next 🚧

Pretrained Models Security

Manage and secure pretrained AI models from external sources with verification, scanning, and safe deployment practices.

Why Pretrained Model Security Matters​

Pretrained models from external sources carry risks:

  • Unknown provenance - Origin may be uncertain
  • Hidden backdoors - Malicious triggers embedded
  • Vulnerabilities - Known security issues
  • License compliance - Legal requirements
  • Data leakage - Sensitive training data exposure

Pretrained Model Features​

Model Registry​

Central repository for pretrained models:

  • Model metadata storage
  • Version management
  • Access control
  • Usage tracking
  • Security status

Provenance Verification​

Verify model origins:

  • Source validation
  • Creator verification
  • Chain of custody
  • Cryptographic signatures
  • Authenticity checks

Security Scanning​

Comprehensive model analysis:

  • Malware detection
  • Backdoor scanning
  • Vulnerability assessment
  • Weight analysis
  • Serialization security

License Management​

Track model licenses:

  • License type identification
  • Compliance requirements
  • Usage restrictions
  • Attribution needs
  • Commercial use rights

Using Pretrained Models​

Adding a Model​

  1. Navigate to AI Security → Pretrained Models
  2. Click Add Model
  3. Provide model information:
    • Name and version
    • Source location
    • Model type
    • Intended use
  4. Upload or link model
  5. Initiate security scan

Security Assessment​

Each model undergoes:

  1. Provenance check - Verify source authenticity
  2. Integrity scan - Check for tampering
  3. Malware scan - Detect malicious code
  4. Vulnerability scan - Find known issues
  5. License analysis - Verify compliance

Assessment Results​

Results include:

  • Security Score - Overall safety rating
  • Findings - Detailed issues found
  • Recommendations - Actions to take
  • Approval Status - Safe to use or not

Approving Models​

Models can be:

  • Approved - Safe for production use
  • Conditional - Approved with restrictions
  • Pending - Awaiting review
  • Rejected - Not safe to use

Model Lifecycle​

Onboarding​

  1. Submit model for review
  2. Automatic scanning
  3. Human review (if needed)
  4. Approval decision
  5. Registration in catalog

Maintenance​

  • Regular re-scanning
  • Vulnerability updates
  • Version management
  • Usage monitoring
  • Compliance tracking

Retirement​

  • Deprecation notice
  • Usage migration
  • Access revocation
  • Archive or deletion
  • Documentation update

Safe Deployment​

Pre-deployment Checklist​

  • Model approved in registry
  • Latest security scan passed
  • License compliance verified
  • Access controls configured
  • Monitoring enabled

Deployment Controls​

  • Approved models only
  • Version pinning
  • Rollback capability
  • Access logging
  • Performance monitoring

Post-deployment​

  • Continuous monitoring
  • Periodic re-assessment
  • Incident response plan
  • Update procedures

Model Categories​

By Source​

  • Open Source - Community models
  • Commercial - Vendor-provided
  • Internal - Organization-trained
  • Research - Academic sources

By Risk Level​

  • Low Risk - Well-established, widely used
  • Medium Risk - Newer or less verified
  • High Risk - Unknown source or issues found
  • Critical Risk - Known vulnerabilities

Best Practices​

  1. Never use unscanned models - Always verify before use
  2. Maintain a central registry - Track all models
  3. Regular re-assessment - Security status changes
  4. Document everything - Source, purpose, approvals
  5. Monitor model behavior - Detect anomalies in production

Related: