Skip to main content

Security

This guide outlines the built-in security, authentication, authorization, and compliance features engineered within MatrixHub to safeguard enterprise AI model assets.

🔒 1. Multi-Tenant Project Isolation​

MatrixHub segregates all resources into logical boundaries called Workspace Projects.

  • Encapsulated Credentials: API tokens, user memberships, and repository permissions are scoped strictly within the project. A token belonging to project-a cannot download or even query model metadata stored inside project-b.
  • Network Segregation: Allows mapping local GPU clusters to specific Projects, ensuring compute nodes only access approved weights.

👤 2. Authentication & Identity Providers (SSO)​

We integrate seamlessly with standard enterprise user directories to ensure unified credential management:

  • LDAP / Active Directory: Authenticate engineers and administrators using their standard corporate user directory.
  • OIDC / OAuth 2.0: Integrate single-sign-on (SSO) providers (e.g. Okta, Keycloak, Ping Identity) to manage dashboard access securely.

🛡ī¸ 3. Fine-Grained Role-Based Access Control (RBAC)​

Access control inside projects is managed through three predefined functional roles:

Project RoleAllowed ActionsTypical Assignment
OwnerFull admin control, invite/delete members, delete repositories, configure replication sync links.Platform Engineers, Devops Leads
ManagerUpload weights, download models, edit repository descriptions, commit and toggle tag locks.ML Engineers, Algorithm Researchers
ReporterRead-only access to query metadata and download cached/locked model weights.Automated GPU Compute Nodes, CI pipelines

📋 4. Compliance Audits & Trail Logging​

To satisfy strict regulatory audits (SOC2, ISO 27001, financial guidelines), MatrixHub records an immutable audit trail for all system-level and API-level events.

Every log entry captures:

  • Who: The authenticated user or API token ID.
  • What: The exact action (e.g., model.upload, tag.lock, token.create).
  • Where: Client IP address and geographical metadata.
  • When: Highly accurate cryptographic timestamp.

Audit logs cannot be modified or deleted by project Managers or Reporters, providing reliable forensic trails.

đŸĻ  5. Malware Scanning & Code Integrity Checks​

AI models (especially standard PyTorch pickle serializations) can act as arbitrary code execution vectors.

MatrixHub integrates active vulnerability scanning:

  • Malicious Code Scan: Automatically scans uploaded weights (Safetensors and pickle checkpoints) upon upload to detect malicious system calls or payloads.
  • Cryptographic Model Signing: Generates cryptographic signatures for approved weights. Compute servers automatically verify these signatures upon fetching, rejecting any tampered weights.