Attacking and Defending Machine Learning Applications of Public Cloud
summary
Claim: a two-step black‑box attack builds a tiny labeled set (≈200 images, ≤2 queries/image) and trains a substitution model to produce transferable adversarial examples that break major cloud classifiers. The method uses a multi‑layer feature loss called FFL‑PGD on a ResNet‑152 substitute and achieves >90% success against Amazon, Google, Microsoft and Collaborate while traditional ensemble baselines stay <50%. Defenses tested: input reconstruction, full adversarial training, and open source tooling AdvBox and Perceptron Benchmark 1.0.
AI-written summary · about
chapters (4)
related talks
36mAdversarial Machine Learning: How Secure Is ML?
21mAdversarial Attacks
20mMachine Learning Attacks – A New Era of Security Epidemic
36mAdversarial Attacks on Deep Learning
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
22mWith Great Training Comes Great Vulnerability: Practical Attacks against Transfer Learning
28mSecurity of Machine Learning
34mHacking the Blackbox