Hacking and Securing ML Systems and Environments
summary
Shows how a malicious prebuilt model or container can open a port (example: port 4000) or drop a miner, letting an attacker pivot from a public notebook instance into private DBs; training jobs and inference endpoints are common entry points. Recommends network isolation to block exfiltration, automated image and host scans with Amazon Inspector on every push, CI/CD manual approvals, IAM least-privilege, and templated deployments via infrastructure-as-code to prevent thousands of exploitable vulnerabilities and enable repeatable remediation.
AI-written summary · about
chapters (16)
- 0:00ML security introduction
- 1:10ML process overview
- 6:23Security testing essentials
- 12:24Securing ML mindset
- 14:51Attacking ML infrastructure
- 16:39Securing ML tools
- 19:54Pre-built model risks
- 23:42Secure ML deployment
- 25:50ML pipeline protection
- 27:35Network isolation training
- 30:48ML vulnerability management
- 32:57Infrastructure as code
- 35:25Monitoring ML environments
- 37:52Security team training
- 39:17Cost versus security
- 42:27Summary
related talks
37mBeyond the Basics: Hacking and Securing Machine Learning Systems and Environments
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
50mBuilding AI Security In: MLSecOps in Practice
26mRunning Secure AI/ML Experiment in a Unified MLOPS-DevSecOps Pipeline
53mSecurity of Machine Learning Systems
42mMLOps Under Attack: Threat Modeling Modern AI Systems
21mEnd-to-end secure ML Development
35mGenAI Application Security: Not Just Prompt Injection