Feature/dynamic backdoor gan

[prachi0606]

#Overview

I am an MSc Data Science student at the University of Edinburgh, contributing this feature as part of my research on adversarial machine learning. My work focuses on designing and evaluating GAN-based dynamic backdoor attacks, and this PR integrates my implementation into ART to support the broader adversarial ML community.

Description

This PR introduces a GAN-based Dynamic Backdoor poisoning attack to the Adversarial Robustness Toolbox (ART).

• Summary: Adds a new attack class DynamicBackdoorGAN that generates input-specific, adaptive perturbations instead of static patches.

• Motivation: Traditional backdoor implementations in ART use fixed triggers, which are easier to detect. This contribution provides a modern, stealthy, and research-driven attack for benchmarking model robustness.

• Files added/updated:

  • art/attacks/poisoning/dynamic_backdoor_gan.py – implementation of the new attack
  • art/attacks/poisoning/__init__.py – registered the attack
  • examples/dynamicbackdoorgan_demo.py – example usage on MNIST and CIFAR-10
  • (optional) docs/poisoning/dynamic_backdoor_gan.md – usage documentation

Fixes # (no open issue, new feature contribution)

---

Type of change

• New feature (non-breaking)
• Improvement (non-breaking)
• Bug fix (non-breaking)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

---

Testing

• Verified attack on MNIST and CIFAR-10 datasets with ResNet18, ResNet50, DenseNet121 and MobileNetV2 models.
• Evaluated clean accuracy, poisoned accuracy, and Attack Success Rate (ASR).
• Confirmed reproducibility via examples/dynamicbackdoorgan_demo.py.

Test Configuration:

• OS: Ubuntu 22.04 (Colab/Local)
• Python: 3.10
• ART: 1.20 (current main)
• Frameworks: PyTorch 2.2, TorchVision 0.17

---

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests/examples proving that the attack works
• New and existing unit tests/examples pass locally with my changes
• My changes have been tested using both CPU and GPU devices

[beat-buesser]

@prachi0606 Do you need any help with this PR?