AUSLab-AdversarialML

Adversarial attack testing for the VPNet model.

Overview

This project provides a framework for evaluating the robustness of the VPNet model against adversarial attacks. It implements various attack methods and measures their effectiveness on ECG data.

Installation

required dependencies: pip install -r requirements.txt

Usage

Running Attacks

You can run adversarial attacks using the provided Python interface. The main entry point is the AttackPack.main() function, which accepts several parameters to configure the attack.

Example

import AttackPack

AttackPack.main(
    method="fgsm",                # Attack method name (e.g., "fgsm", "pgd", etc.)
    eps=[0.1, 0.01, 0.001],       # List of epsilon values for the attack
    device="cpu",                 # Device to run the attack on ("cpu" or "cuda")
    weights="model_weights.pt",   # Path to the model weights file
    adv_method="d",               # Attack type: "d" (data), "r" (rr), or "dr" (both)
    early_stop=10                 # (Optional) Stop after N successful attacks
)

Parameters

• method: Name of the attack method to use. Multiple methods can be specified as a comma-separated string.
• eps: List of epsilon values (perturbation strengths) to test.
• device: Device to run the computations on ("cpu" or "cuda").
• weights: Path to the trained VPNet model weights file.
• adv_method: Type of adversarial attack:
  • "d": Perturb data only
  • "r": Perturb rr only
  • "dr": Perturb both data and rr
• early_stop: (Optional) Integer. If set, stops after the specified number of successful attacks.

References

• VPNet Paper/Repository (add link if available)