search

cleverhans-lab / cleverhans

An adversarial example library for constructing attacks, building defenses, and benchmarking both
MIT License
6.21k stars 1.39k forks source link

Color Images Adverserial distortions #391

Closed aashish-kumar closed 6 years ago

aashish-kumar commented 6 years ago

Hello, I tried using the attacks to generate adversarial for Color Images but the attacks introduce color artifacts in the Images. Has there been work done toward generating adversarial without significant color artifacts? Thanks

gokul-uf commented 6 years ago

Could you give an example? I've seen this happen when my epsilon was too large.

On Wed 18 Apr, 2018, 2:03 PM aashish-kumar, notifications@github.com wrote:

Hello, I tried using the attacks to generate adversarial for Color Images but the attacks introduce color artifacts in the Images. Has there been work done toward generating adversarial without significant color artifacts? Thanks

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tensorflow/cleverhans/issues/391, or mute the thread https://github.com/notifications/unsubscribe-auth/AFJebSdB5Gxx6n8dDaB-GeZdttupn1Dmks5tpyuQgaJpZM4TZ7_0 .

aashish-kumar commented 6 years ago

I tried JSMA and Carlini Wagner with default settings and FGSM with epsilon 0.3 Its working fine for grayscale Images but with Color Images the adverserial Noise is too apparent

AlexeyKurakin commented 6 years ago

Could you show an example of color artifacts?

Sometimes color artifacts happen when you forget to clip an image values and then tried to save or show it. For example, by default matplotlib assumes that all pixels of float32 image are in range [0; 1]. If any values are outside of this range then image will have weird color artifacts when shown. If this is your problem then the solution is to clip values of all channels of all pixels to be in [0, 1] range.

goodfeli commented 6 years ago

for images in [0, 1] epsilon of 0.3 should definitely be apparent. If you want it to be totally imperceptible eps needs to be less than 1/255

aashish-kumar commented 6 years ago

@All The theta for gamma was set to 1.0 for JSMA attack which was resulting in uneven color changes in the adversarial Images. reducing it to 0.1 produced good attacks. Thanks