Security of Machine Learning
summary
Overview of concrete risks and tools: poisoning corrupts training data to inject backdoors while evasion crafts FGSM or Carlini‑Wagner perturbations at deployment to make real-world objects (glasses, a stop sign, a 3D-printed turtle) be misclassified—turtle becomes “rifle,” stop sign read as “go.” Public libraries like Cleverhans and IBM toolkits reproduce attacks; defenses (data sanitization, reject on negative impact, defensive distillation, dimensionality reduction, distance‑to‑boundary membership scores) exist but recent evaluations show each can be bypassed.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 0:18Background
- 1:11Security of Machine Learning
- 2:21Why do we need secure machine learning
- 3:16Main goal of this talk
- 3:33Deployment of ML
- 4:33Confidentiality Integrity Availability
- 6:43Poisoning
- 7:41Google Translate
- 8:05Evasion Attack
- 9:00Adversary Attack
- 9:45Turtles
- 10:44Stop signs
- 13:18Attack Libraries
- 14:04Transferability
- 17:21Wrong
- 21:26Other vulnerabilities
- 22:09Summary
- 22:41Open problems
- 23:36Questions
related talks
53mSecurity of Machine Learning Systems
36mAdversarial Machine Learning: How Secure Is ML?
20mMachine Learning Attacks – A New Era of Security Epidemic
44mComputers are Stupid: Protecting "AI" from Itself
46mA Short Introduction to Hacking Artificial Intelligence
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence