Skip to content

Latest commit

 

History

History
15 lines (10 loc) · 562 Bytes

README.md

File metadata and controls

15 lines (10 loc) · 562 Bytes

Creating adversarial images to fool an MNIST classifier

Python 3 implementation of Carlini & Wagner's algorithm (https://arxiv.org/abs/1608.04644).

Adversarial Attack
The parameters and variables in Adversarial-Attack.ipynb follow the notation in the paper.

Distance Metric
L2 distance between x and x' is computed ||x - x'||2.

Objective Function
Find delta that solves: min ||delta||2 + c * f(x + delta), such that x+delta is a subset of [0,1]^n.

Box Constraint
delta_i = 1/2 (tanh(w_i)+1) - x_i.