What is the main security risk associated with downloading pre-trained models from repositories without verifying provenance?
- A.A. The models may have lower accuracy than advertised, leading to poor business outcomes
- B.C. The models may be too large for production deployment, causing infrastructure costs
- C.B. The models may contain backdoors, adversarially perturbed weights, or pickle-based malicious code that executes upon loading
- D.D. The models may not be licensed for commercial use, creating legal risk
Why C is correct
Unverified pre-trained models from public repositories represent an ML supply chain risk: they may contain backdoor triggers, subtly modified weights that degrade accuracy on specific inputs, or malicious serialization code (in pickle-format files) that executes arbitrary commands when the model is loaded. MITRE ATLAS catalogs this as ML Supply Chain Compromise. Accuracy, infrastructure, and licensing are separate concerns.
Know someone studying for AI Security Fundamentals? Send them this one.