Limitations for features

[Arkadiusz Warzyński]

Hi everyone,

this can be a trivial problem for many of you, but how can I define the space of values for a given feature? I have a model that contains many binary features but adversarial examples created by FSGM algorithm have a values like '0.7' or '0.3' assigned to this features. Is there a way for the algorithm to generate values only from specified range?

[Nicolas Papernot]

There are several ways you could go about that.

If you're looking at binary features, you can for instance follow the sign of the gradient to decide which features to set to 1 or 0. That would require making small changes to the FGSM.

You can also look at how we adapted the jsma for a similar setting here: http://arxiv.org/abs/1606.04435

[Arkadiusz Warzyński]

I am working on anomaly detection for a network traffic and I want to prove if this kind of network security systems have a vulnerability to adversarial examples. The article, which you have linked, presents a similar problem so it should be very helpful. 

I am unfortunately still not so good Python and TensorFlow programmer, so I will write here when I have a problem with implementation.

[Arkadiusz Warzyński]

I have read the article and I see one difference. My model contains both binary and numerical values (also categorical, but i use one-hot). How can I define different solutions for different types of feature?