Skip to content

Latest commit

 

History

History
 
 

defense_jump

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 

Jump Activation function

This defense changes the activation function to a new discontinuous Jump function.

Defense Idea

Some say that adversarial examples exist because classifiers are piecewise continuous, and therefore must interpolate between regions that don't make sense visually. A discontinuous function would be able to better model the space and prevent adversarial examples. The Jump activation function is like ReLU, but has a discontinuity at zero:

jump(x) = 0 if x < 0, otherwise 5 + x

Classifiers trained with this jump function appear robust.

Training

We train the model exactly as normal but with this new jump function. In order to make training converge, we increase the jump constant from 0 up to 5 throughout the course of training.

Objectives

Our best attack results are as follows, when evaluated on the first 100 images in the CIFAR-10 test set using the provided pretrained model..

  • l_infinity distortion of 4/255: 92% attack succes rate

References

Defenses

This defense idea is taken from an unpublished manuscript by Ian Goodfellow developed in 2018.

Attacks

No published attacks on this particular activation function exist.