v0.6.0

What's New

• PNAPatchOut - Towards Transferable Adversarial Attacks on Vision Transformers
• TGR - Transferable Adversarial Attacks on Vision Transformers with Token Gradient Regularization

Full Changelog: v0.5.0...v0.6.0

v0.5.0

What's Changed

• The SSP (self-supervised) attack by @spencerwooo in #11

Full Changelog: v0.4.0...v0.5.0

v0.4.0

What's New

• Add support for Python 3.12 with PyTorch 2.2.
• Add FIA attack support.

Full Changelog: v0.3.0...v0.4.0

v0.3.0

🚧 Breaking Changes

1. The name of parameter transform has been renamed to normalize

2. The order of function parameters have changed:

-FGSM(model, transform, **attack_cfg, device)
+FGSM(model, normalize, device, **attack_cfg)

v0.2.2

• New attack torchattack.Admix implemented with reference from TensorFlow version (JHL-HUST/Admix).
• requirements.txt removed. To develop locally, install the library with pip install -e '.[dev]'.
• Formatter black replaced with ruff format.

v0.2.1

Default alpha for gradient-based attacks are fixed. (Set to eps / steps if not specified.)

Full Changelog: v0.2.0...v0.2.1

v0.2.0

Bumped support for Python 3.11.

requires-python = ">=3.10"

v0.1.4

What's Changed

• New attack: DeepFool, torchattack.DeepFool.

Full Changelog: v0.1.3...v0.1.4

v0.1.3

What's Changes

• Ship package with py.typed to make type hints available when installed.
• Additional comments for module torchattack.dataset.

Full Changelog: v0.1.2...v0.1.3

v0.1.2

What's Changed

• Update dependencies black, ruff, pre-commit, and mypy
• Remove an additional line (bug) in VNI-FGSM

Full Changelog: v0.1.1...v0.1.2