CGammaSectionsEvasionProblem attack budget

[pralab]

CGammaSectionsEvasionProblem attack budget

[zangobot]

To better understand the attack, I suggest you to have a read at the paper of the attack: https://arxiv.org/abs/2003.13526
The "iterations" parameters specify the number of iteration, while the population size specifies how many variants of the same sample are generated at each round (see the genetic algorithm explanation in the paper).

The attack injects content from goodware samples (that you specify with the "goodware_folder" parameter).
The "how_many" parameter is used to decide how many goodware extract from that folder.

Regarding Android malware, no, this repository is only intended to manipulate Windows malware.
Attacks against those detectors are provided to work only in feature space in the original SecML library (but there are no manipulations attached).
But, if you would like to extend it and create input-space attacks on Android, you can open a pull request with your work!

And thank you for citing my work, I hope it was helpful so far!

[zangobot]

Hello, the evasion depends on the threshold that you have. Here the optimizer stops after 11 queries (because it is likely that it found a region where the decision function is flat).
Try to increase the number of sections used for the attack, and try to decrease the regularization parameter.
The success of the attack also depends on the goodware you are using to inject content.

[QuangNguyen2609]

u mean the section used for attack is "how many" parameter right ? and I'm not sure what's the effect of population size to this attack and how many goodware is enough for this attack (I have 7000 malware). Im also wondering if goodwares need to be specific to run the attack successfully

thanks alot

[QuangNguyen2609]

Ah one more thing, im wondering are there any attacks that i can run for Android Malware Detection with Drebin like the tutorial ? It would be helpful if you could show me how to initialize those attacks. I will cite your work after my paper is done

[quangnguyen-uoa]

Thakns you very much. I have one last question as I follow the CAttackEvasionPGDLS attack for Android Malware Detection. I ran the debugger several time but I haven't found the contribution of dmax (attack budget). Can you show it to me and what does the dmax do to the attack please ? Thanks alot

[zangobot]

Wait wait that is another library 😄
Please open an issue there to get help with the SecML library (also, I don't think I understood the question).
I am closing this issue for now.