Attacking And Protecting Artificial Intelligence
summary
Launches an OWASP AI security and privacy guide and points to ISO 5338 and dataset/versioning as concrete standards to adopt. Demonstrates real attacks: data poisoning via traffic‑sign stickers, Tay/Bing model extraction and a membership inference revealing patient status, plus model theft from repeated API queries. Recommends concrete defenses: hide weights, add noise to training, throttle API calls, enforce dataset access controls, unit‑test data pipelines and treat production models as maintainable software.
AI-written summary · about
chapters (14)
- 0:00Introduction to AI risks
- 1:33Speaker background and guides
- 5:06AI application roles
- 7:29Understanding AI types
- 11:46AI engineering practices
- 18:01Production AI as software
- 20:57AI attack surface
- 23:22Data poisoning threats
- 28:18Model manipulation
- 30:57Model inversion and theft
- 37:14Privacy and transparency
- 40:27Data governance and fairness
- 44:38Ethical AI regulations
- 47:46Q&A and conclusion
related talks
57mAI Applications Security Puzzle
31mSecurity Considerations for Services Using AI Models
17mHands-on Practice
53mHow AI is impacting cybersecurity. Does it?
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
49mThe Intersection of AI and Cybersecurity
31mSecuring the Future: Navigating AI Vulnerabilities and Evolving Security Practices
45mThreat Modeling in the Age of AI