PHWSODownload
Established
AI Supply Chain Security Models, Datasets, Dependencies
Published 9/2026
MP4 | Video: h264, 1920x1080 | Audio: AAC, 44.1 KHz, 2 Ch
Language: English | Duration: 11h 23m | Size: 4.32 GB
Secure models, datasets and ML dependencies: provenance, signing, scanning, SBOM/AI-BOM and pipeline hardening
What you'll learn
Map your AI supply chain end to end: assets, actors, trust boundaries and invisible dependencies
Inspect model files safely and set a serialization policy across pickle, safetensors, Keras, GGUF and ONNX
Build a model intake pipeline that scans, sandboxes and converts untrusted models before production
Verify provenance using Sigstore, OpenSSF model signing, SLSA build provenance and in-toto attestations
Produce an AI-BOM and evidence EU AI Act, Cyber Resilience Act and NIST SP 800-218A supply chain duties
Requirements
Comfort with the command line, Python packaging and CI/CD pipelines
No machine-learning research background required; familiarity with how models are deployed is enough
Description
"This course contains the use of artificial intelligence."
Your models come from somewhere. So do your datasets, your adapters, your tokenizers, your GPU runtime and the dozens of Python packages underneath them. This course treats that whole path, from a public model hub to a running inference server, as a supply chain, and shows you how to secure it.
We start with a working threat model: ingest, build, serve. You will map your own trust boundaries, inventory the assets and the actors you depend on, and find the places where provenance quietly breaks in a normal ML team. From there the course goes format by format, so you understand why a model file is not like other files: how Python pickle executes code on load, what a PyTorch checkpoint actually contains, what safetensors fixes and what it does not, and how Keras archives and edge inference formats widen the parser attack surface.
Then we build the controls. You will design a model intake pipeline that scans, sandboxes and converts untrusted artifacts; verify provenance with Sigstore, OpenSSF model signing, SLSA build provenance and in-toto attestations; harden dependency resolution against confusion, typosquatting and slopsquatting; lock down CI/CD, registries and promotion gates; and reduce the blast radius at runtime with egress controls and workload isolation.
The last third turns controls into a programme: dataset provenance records, agent and MCP tool allowlists, SBOM and AI-BOM production, acceptance policy, release gates, and the regulation that binds you, including EU AI Act Articles 15 and 25, the Cyber Resilience Act and NIST SP 800-218A.
Every lecture teaches a control and then shows the incident where its absence cost someone. The course is vendor-neutral: tools are named as examples, never ranked. Where a control is imperfect we say so, because scanners are a filter and not a gate, and build provenance proves where a build ran and not that it was clean. Labs and checklists run throughout so you can apply each step to your own stack.
Who this course is for
Security engineers, application security teams, MLOps and platform engineers deploying third-party AI models
Technical risk and compliance staff who must evidence AI supply chain controls to auditors or regulators
Homepage
Code:
https://www.udemy.com/course/ai-supply-chain-security-models-datasets-dependencies/DOWNLOAD LINKS
