Tensorflow, there is no spoon
summary
Shows how to craft adversarial images that flip labels on real models: the 2016 Google example turns a panda into a gibbon at 99.3% confidence, AlexNet gets many images labeled ostrich, and a 3D‑printed perturbed turtle is classified as a rifle. Uses gradient steps against Inception v3 to push a cat image until the model is 98% confident it’s a spoon, and reproduces physical attacks with 3D-printed adversarial glasses. Defense is practical: augment training with adversarial examples or use GANs to harden models.
AI-written summary · about
chapters (10)
- 0:00Start
- 11:51How Neural Network Works
- 12:28How a Neural Network Works
- 13:54Calculation for the First Neuron in the First Hand Layer
- 21:33Inception V3
- 22:27Guinea Pig
- 24:49Gradient Descent
- 26:59Results
- 30:18Protect Your Model against Attacks
- 42:54Optical Illusions
related talks
45mTensorflow, there is no spoon
38mTensorflow, there is no spoon
38mUsing adversarial samples to robustify your Neural Network Models
25mTricking Neural Networks : Explore Adversarial Attacks
27mHow to Trick a Neural Network
36mAdversarial Attacks on Deep Learning
44mComputers are Stupid: Protecting "AI" from Itself
21mAdversarial Attacks