Strategy Network Architecture

[sundasfatema]

Hello, I hope you're well. I went through your paper and supplementary material to understand how a strategy network generates attack types. My understanding through code and architectural diagram is as follows:
The strategy network generates a range of attack types by configuring the following parameters:

Attack Type: The specific method used to generate adversarial examples.
Epsilon: The magnitude of perturbation added to the inputs.
Number of Iterations: How many times the attack algorithm is applied.
Step Size: The size of each step in iterative attacks.
By sampling different combinations of these parameters, the strategy network can produce a diverse set of attack strategies, each potentially effective in different ways or against different models.

While going through code, I could understand everything else but the variable args.attack_types. Basically I want to understand what are the attack types that this strategy network caters. What are the types of attack that this architecture covers for instance PGD, FGSM? I cannot find that information in code/paper. Kindly help.

[sundasfatema]

Hello, I wanted to know if you can send me official email so that I can email you since the email in the paper isn't working. I wanted to learn more about strategy network. Thank you.

[jiaxiaojunQAQ]

Hello, I hope you're well. I went through your paper and supplementary material to understand how a strategy network generates attack types. My understanding through code and architectural diagram is as follows: The strategy network generates a range of attack types by configuring the following parameters:

Attack Type: The specific method used to generate adversarial examples. Epsilon: The magnitude of perturbation added to the inputs. Number of Iterations: How many times the attack algorithm is applied. Step Size: The size of each step in iterative attacks. By sampling different combinations of these parameters, the strategy network can produce a diverse set of attack strategies, each potentially effective in different ways or against different models.

While going through code, I could understand everything else but the variable args.attack_types. Basically I want to understand what are the attack types that this strategy network caters. What are the types of attack that this architecture covers for instance PGD, FGSM? I cannot find that information in code/paper. Kindly help.

Thank you for your attention to our work. We apologize for the late response. The args.attack_types in the code from the paper does not have any specific meaning. In this work, we use PGD (Projected Gradient Descent) attacks as the baseline to find the optimal adversarial noise.

[jiaxiaojunQAQ]

Hello, I wanted to know if you can send me official email so that I can email you since the email in the paper isn't working. I wanted to learn more about strategy network. Thank you.

My latest Email is [email protected]. The strategy network uses ResNet18 as the backbone, with multiple independent fully connected layers （Refer to https://github.com/jiaxiaojunQAQ/LAS-AT/blob/master/LAS_PGD_AT/StrategyNet.py）.