Secure AI: Integrating OWASP Principles in Machine Learning Model Development
summary
Survey of OWASP Top10 for ML maps ten real risks and pragmatic fixes: adversarial patches that flip a French horn into a “hot dog” or a stop sign into a speed limit, data poisoning from public Azure blobs (one gov example), transfer‑learning weight tampering, model theft and membership inference. Recommended practices are concrete: adversarial training and input validation, model ensemble and anomaly detection, RBAC and private blob storage, Dockerized CI/CD with versioned model registry, monitoring/logging and regular retraining.
AI-written summary · about
chapters (11)
- 0:00Secure AI introduction
- 3:18Why ML needs security
- 10:10OWASP ML Top 10
- 16:56ML system vulnerabilities
- 23:50Data poisoning & storage
- 30:17Model inversion attacks
- 37:02Transfer learning risks
- 43:00ML pipeline security
- 47:19Traditional vs ML security
- 53:20Secure MLOps deployment
- 59:41Q&A
related talks
50mBuilding AI Security In: MLSecOps in Practice
49mSecuring AI Apps with the OWASP Top Ten for Large Language Models
45mThreat Modeling in the Age of AI
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
28mSecurity of Machine Learning
31mSecurity Considerations for Services Using AI Models
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
53mSecurity of Machine Learning Systems