Secure ML: Automated Security Best Practices in Machine Learning
summary
Trains a scikit‑learn logistic‑regression on the iris data set, pickles and deploys it as a Kubernetes microservice via seldon core deployment, then demonstrates a malicious pickle that runs arbitrary code (creates a pawn txt) so the exploit actually works. Highlights adversarial-image attacks and shows detectors like alibi detect, dependency and image‑scan failures, and recommends reproducible environments and package locking to prevent supply‑chain surprises. Concludes with concrete CI/CD and runtime checks.
AI-written summary · about
chapters (19)
- 0:00Intro
- 1:28Security Challenges in Machine Learning
- 3:38Standards - OWASP
- 4:21TLDR; Machine Learning Deployed
- 7:19Security Areas in Phases of MLOPS
- 8:27Security Risk - Artifacts
- 10:44Artifacts & Pickles
- 11:16Artifact Trust or Discard
- 12:56Access to the model (1/3)
- 14:59Access to the model (2/3)
- 16:23Dependencies
- 18:14Dependency-Chain Risks
- 18:47Security Risk 3 - Live Dependency Install
- 19:30Security Risk 3 - Dependency Scans
- 19:55Codescan
- 20:33Jupyter Notebooks
- 20:43Images
- 21:00Industry Standards: Flawed ML Top 10
- 22:01Resources for Talk
related talks
21mEnd-to-end secure ML Development
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
53mSecurity of Machine Learning Systems
37mBeyond the Basics: Hacking and Securing Machine Learning Systems and Environments
43mHacking and Securing ML Systems and Environments
40mMLSecOps with Automated Online and Offline ML Model Evaluations on Kubernetes
28mSecurity of Machine Learning
32mPoisoned Pickles Make You Ill