Overview

I am a Master of Science student in Data Science at the University of Edinburgh, focusing on advanced analytics, machine learning, and data-driven insights.

Description

This PR adds a new poisoning attack to ART: DynamicBackdoorGAN, which generates dynamic, input-specific backdoor triggers using a lightweight CNN generator. Unlike static patch-based attacks, this approach produces adaptive perturbations tailored to each input, making the attack more stealthy and powerful.

Motivation:
To introduce a more realistic and effective backdoor attack strategy aligned with recent research into dynamic triggers using GAN-like architectures.

Fixes: N/A (New feature, no associated GitHub issue)

Type of change

Please check all relevant options.

• New feature (non-breaking)

Testing

The implementation was verified locally by training multiple model (ResNet18, ResNet50, MobileNetV2, DenseNet121) classifiers on CIFAR-10 and MNIST datasets using:

• Clean training accuracy
• Poisoned training accuracy
• Attack Success Rate (ASR)

A minimal unit test was added under tests/attacks/poisoning/backdoor/ to validate:

• Poisoning logic
• Trigger application
• Input-output shape consistency

Test Configuration:

• OS: Ubuntu 22.04 (Colab environment)
• Python version: 3.10
• ART version: Latest (cloned from GitHub)
• PyTorch version: 2.1.0

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
• I have made corresponding changes to the documentation (inline docstrings)
• My changes generate no new warnings
• I have added tests that prove my feature works
• New and existing unit tests pass locally with my changes
• My changes have been tested using both CPU and GPU devices