Security & Privacy
AI Software Supply Chain Security
Category
Security & Privacy
Abstraction level
System
Operation level
DeploymentData
Use cases
Safely downloading models from registriesProtecting MLOps pipelines from poisoningCompliance and security auditing of AI systems
How it works
Practices include provenance and signature verification of artifacts, scanning models and packages for malicious code (e.g. unsafe pickle files), weight-integrity checks, an AI SBOM, environment isolation and trusted-source policies for models.
Problem solved
How to ensure that the models, data and dependencies used in an AI system have not been tampered with, poisoned or substituted along the way?
Strengths & limitations
Strengths
✓Covers AI-specific attack vectors (poisoning, weight backdoors)
✓Builds trust in external models and data
✓Supports regulatory compliance
Limitations
✗Hard to verify model weights for hidden backdoors
✗Broad and complex supply chain
✗Lack of mature, widely adopted standards