Flawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
summary
Model serving surfaces expose concrete failures: pickle deserialization lets a loaded model run arbitrary system calls, pip installers can execute code at install time (typosquatting risks), and public inference endpoints enable adversarial inputs and extraction of training data (LLMs leaking secrets). Mitigations used here are artifact signing, dependency scanning and lockfiles (Dependabot, Bandit, Trivy/Snyk), runtime payload detectors (Alibi Detect), and deploying models from prebuilt model environments with RBAC and mTLS enforced.
AI-written summary · about
chapters (11)
- 0:00ML security challenges
- 4:29Serving stage in ML
- 8:21Model serialization risks
- 11:55Pickle-based vulnerabilities
- 13:39Adversarial attacks on models
- 16:05Custom inference code risks
- 18:13Dependency risks in ML
- 21:41Securing Docker images
- 24:28OWASP-like top 10 for ML
- 26:51Secure ML platforms
- 27:55ML SecOps resources
related talks
28mSecurity of Machine Learning
43mHacking and Securing ML Systems and Environments
21mEnd-to-end secure ML Development
37mBeyond the Basics: Hacking and Securing Machine Learning Systems and Environments
35mGenAI Application Security: Not Just Prompt Injection
50mBuilding AI Security In: MLSecOps in Practice
43mLogically Securing the Illogically Logical Use of Large Language Models
25mSecure ML: Automated Security Best Practices in Machine Learning