How to trick a neural network!
summary
Demonstrates that a Google‑trained image classifier can be nudged from “giant panda” to “vulture” by computing gradients and applying many tiny pixel perturbations, using a prebuilt GoogleNet model (50 megabyte download), standard backprop/derivative calls and Docker on a CPU, running about 100 tiny steps that take 8–10 minutes on a laptop, and producing near‑certain misclassification rather than mere visual examples.
AI-written summary · about
chapters (8)
- 0:00Introduction to confusion
- 1:37Discovering adversarial attacks
- 2:32Understanding how models work
- 4:04Setting up the environment
- 5:50The mathematics of trickery
- 8:14Implementing the attack
- 9:48Panda to vulture analysis
- 11:09Conclusion and resources
related talks
27mHow to Trick a Neural Network
34mNeural networks: insane in the membrane
43mTensorflow, there is no spoon
41mDeep Learning to Solve Challenging Problems
25mTricking Neural Networks : Explore Adversarial Attacks
53mML BuzzWords demystified
48mKeynote – Sameer Singh
38mUsing adversarial samples to robustify your Neural Network Models