Secure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
summary
Timeline: deep nets unleashed in 2012 and real attacks appear by 2016 — Tay taken down after 16 hours is cited as a poisoning example and UW researchers show stop sign stickers fooled model at inference. Defenses presented include adversarial training and defensive distillation, privacy fixes like teacher–student ensembles and sharding/slicing “machine unlearning”, and tooling: CleverHans, Adversarial Robustness Toolbox and LIME. Emphasis on needing MLOps pipelines before applying these costly techniques.
AI-written summary · about
chapters (20)
- 0:00Introduction to AI security
- 1:04Machine learning in daily life
- 2:27What makes a secure system
- 4:53Adversarial ML challenges
- 5:56Training attacks and defenses
- 9:27Inference attacks and real impact
- 10:30Autonomous vehicle vulnerabilities
- 12:35Adversarial training techniques
- 15:02Defensive distillation explained
- 16:25MLOps necessity for security
- 17:27Data ethics and privacy
- 19:11GDPR and model compliance
- 20:16Private aggregation solutions
- 22:45Machine unlearning methods
- 24:27Algorithmic bias risks
- 26:33Compass recidivism bias case
- 29:00Model explainability tools
- 30:02Security software resources
- 31:26Recommended learning resources
- 32:53Q&A and closing remarks
related talks
46mA Short Introduction to Hacking Artificial Intelligence
28mSecurity of Machine Learning
53mSecurity of Machine Learning Systems
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
36mAdversarial Machine Learning: How Secure Is ML?
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications