Tensorflow, there is no spoon
summary
Shows how targeted pixel perturbations and gradients turn a cat photo into a spoon on Inception v3 98.65% — generated by iterative gradient steps taking about three hours on a laptop — and how physical attacks work: 3D-printed glasses create both dodging and impersonation in face-recognition systems and a perturbed turtle prints as a rifle. Offers a practical defence: train with adversarial examples using GAN adversarial training, which raises robustness to fake inputs and even improves clean-image accuracy.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:08Teaser
- 7:58Dodging Attack
- 9:033d Printed Turtles
- 11:20Neural Network
- 12:32Hidden Layers
- 15:47Decision Boundary
- 16:35Adding Hidden Layers
- 19:25Top 5 Error Rate
- 20:53Inception V3
- 21:40Guinea Pig
- 25:57Results
related talks
43mTensorflow, there is no spoon
38mTensorflow, there is no spoon
38mUsing adversarial samples to robustify your Neural Network Models
25mTricking Neural Networks : Explore Adversarial Attacks
44mLearning from Machines
27mHow to Trick a Neural Network
26mNew Face, Who Dis? Recent Adversarial Approaches to Facial Recognition
13mHow to trick a neural network!