Computers are Stupid: Protecting "AI" from Itself
summary
Adversarial examples repeatedly turn real objects into high‑confidence misclassifications (a 3D‑printed turtle gets labeled a rifle at 90% confidence), iterative poisoning lets biometric systems be bypassed by crafted glasses, and membership‑inference attacks recover training membership at 70–90% accuracy. Defenses discussed include feature squeezing input sanitization, locking down model APIs, LIME/Li5 local explanations (not global), and a homomorphic pseudonymization prototype from KIProtect that encrypts or pseudonymizes features while preserving useful relationships for training.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 0:54The Hype
- 3:03Example A Google Translate
- 6:21Selfdriving cars
- 7:59Computers are stupid
- 9:05Adversarial examples
- 11:20Biometrics
- 13:38Malicious intent
- 16:15Bias
- 20:50Privacy
- 29:03Linear regression
- 31:07How to solve this problem
- 32:32Access to API
- 35:22Homomorphic Encryption
- 38:21Interdisciplinary Panels
- 39:39Paper Distribution
related talks
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
28mSecurity of Machine Learning
46mA Short Introduction to Hacking Artificial Intelligence
45mSecuring AI: There Is No Try, Only Do!
20mMachine Learning Attacks – A New Era of Security Epidemic
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
31mSecurity Considerations for Services Using AI Models
01h 05mNavigating the Security and Privacy Landscape of Modern AI