With Great Training Comes Great Vulnerability: Practical Attacks against Transfer Learning
summary
Shows that transfer learning creates a practical black‑box attack that uses a public Teacher model to craft imperceptible perturbations by neuron representation matching at a middle layer, assuming Teacher white‑box and Student black‑box. Tests on self‑trained Students and services (Google Cloud ML, Microsoft CNTK, PyTorch) yield >92% and >95% targeted success and teacher fingerprinting at >88%. A defense adds a neuron‑distance constraint to the transfer objective, cutting attack rates from >90% to 30% and 12% with minimal accuracy loss; code and models are on GitHub.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:27Deep Learning
- 1:22Transfer Learning
- 2:57Deep Learning Basics
- 5:08Student Model Diversity
- 6:09Attack on Deployment
- 6:40Whitebox vs Blackbox
- 7:51New Attack Model
- 8:51Attack Example
- 11:04Testing
- 11:53Fingerpainting
- 13:11Intuition
- 14:54Results
- 15:49Top Questions
related talks
20mAttacking and Defending Machine Learning Applications of Public Cloud
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
25mTricking Neural Networks : Explore Adversarial Attacks
20mMachine Learning Attacks – A New Era of Security Epidemic
21mTricking Neural Networks : Explore Adversarial Attacks
36mAdversarial Attacks on Deep Learning
21mAdversarial Attacks
36mAdversarial Machine Learning: How Secure Is ML?