transferattacks 1.0

A PyTorch library for adversarial attacks

TransferAttacks

Overview

TransferAttacks is a PyTorch library for implementing adversarial attacks with improved transferability. It includes state-of-the-art methods categorized into different approaches:

• Gradient-based: Modify gradients for enhanced transferability.
• Ensemble-based: Use multiple models to improve attack success.
• Generation-based: Train generative models to craft adversarial examples.

Installation

pip install transferattacks

Supported Attacks

Gradient-based Methods

• BPA: Recovers truncated gradients in non-linear layers.
• AGS: Uses adversary-centric contrastive learning.
• MetaSSA: Utilizes low-frequency feature mixing.
• VDC: Adds virtual dense connections for dense gradient backpropagation.
• MA: Minimizes KL divergence between models.

Ensemble-based Methods

• Ens: Generates adversarial examples using multiple models.
• Ghost: Uses dropout and random scaling to create ghost networks.
• SVRE: Uses variance-reduced gradients.
• LGV: Ensembles multiple weight sets.
• MBA: Maximizes average prediction loss using Bayesian optimization.
• AdaEA: Adjusts surrogate model weights dynamically.
• CWA: Exploits common weaknesses in an ensemble.
• SMER: Uses reinforcement learning for weight refinement.

Generation-based Methods

• CDTP: Learns domain-invariant perturbations.
• LTP: Uses mid-level features for perturbation generation.
• ADA: Stochastically perturbs shared salient features.
• GE-ADVGAN: Enhances adversarial transferability via gradient editing.

Usage

from transferattacks import FGSM
attack = FGSM(model, epsilon=0.03)
adv_images = attack.generate(images, labels)

Citation

If you use TransferAttacks in your research, please cite the corresponding papers.

License

MIT License