Implementation of Fast Gradient Sign Attack about secml_malware HOT 6 CLOSED

Hi!

You need to use the FGSM inside the repo for attacking end-to-end models or you need to code a custom attack?
All the attacks are subclass of CAttackEvasion (from secml).
Then, I created other subclasses for the different strategies.
The FGSM specifies only how to perturb the malware and how to compute the gradients and loss, and then it uses the methods of the underlying library.
Most of the details are included in the paper I listed (the "Adversarial EXEmples" paper contains the formalization I used for encoding all the attacks, it might help understanding what does what)

By the way, I still need to write a decent doc, I'll try to do that in the future (time consuming task, but needed)

from secml_malware.

Im trying to integrate the FGSM attack into the tutorial attack to try and generate Adversarial AI Malware.
Im having trouble trying to integrate it as there is no documentation on how to use it.

Could you link me the "Adversarial EXEmples" paper?

from secml_malware.

Hi again!
In the readme of this repository, there is a full list of the paper I used in this topic.
There is also a notebook with an example attack (the partial DOS header attack).

from secml_malware.

Ok, I got the point.
So, the attack needs:

1. the end-to-end model;
2. the list of locations inside the sample that you want to edit;
3. the epsilon parameter

If you want to use default locations, have a look at the CKreukEvasion (from the name of the author).
That attack uses padding and slack space to compute adversarial malware.

from secml_malware.

Hi,
Actually im trying to find out which function from the C_Fast_Gradien_Sign_Evasion.py to use to run the attack.

From the image i am using the apply_feature_mapping() function. Am i using the correct function and what kind of parameters are needed for this function?

from secml_malware.

There is the run function, as shown in the tutorial:

for sample, label in zip(X, y):
    y_pred, adv_score, adv_ds, f_obj = your_attack.run(CArray(sample), CArray(label[1]))

So, after you have created the CFastGradientSignEvasion correctly, you just call that function and you're done.
But, first of all: how do you want to perturb the malware?
Because, as I said before, you could use the CKreukEvasion, as it is a FGSM implementation with standard locations used for perturbing the malware.

from secml_malware.