v.1.7.2

Full Changelog: v1.7.1...v1.7.2

v1.7.1

Fixes

• Fix optional deps still required during import.

Full Changelog: v1.7.0...v1.7.1

v1.7.0

What's Changed

• New attack SGM by @spencerwooo in #41
• New attack BPA by @spencerwooo in #42
• New attack FDA by @spencerwooo in #43
• Dependency import optimizes by @spencerwooo in #40
• Integrate DANAA with ensemble gradient aggregation
• Extend ILPD preprocessing to support additional built-in models

Fixes

• Fix AttackModel preprocessing inference on models using nn.Identity
• Declare optional extra dependencies for some attacks (kornia, scipy) and docs builds

Full Changelog: v1.6.0...v1.7.0

v1.6.0

What's Changed

• Add support for targeted attack evaluation by @spencerwooo in #36
• Implementation of BFA by @spencerwooo in #37
• Implementation of GRA by @spencerwooo in #38
• Implementation of MuMoDIG by @spencerwooo in #39
• FoolingRateMetric has been renamed to FoolingRateMeter

Full Changelog: v1.5.1...v1.6.0

v1.5.1

What's Changed

• Refactor module eval to evaluate by @spencerwooo in #35
• Fix timm models interpolation unresolved

Full Changelog: v1.5.0...v1.5.1

v1.5.0

What's New

• Introduce AttackModel.create_relative_transform and eval.save_image_batch
• Add fooling rate metric eval.FoolingRateMetric reset option
• Update documentation for attack registering

Fixed

• Fix generative arguments not being passed in attack runner eval/runner.py

Refactored

• Refactor attack category checks to use is_category method
• Move attack module to public

Full Changelog: v1.4.0...v1.5.0

v1.4.0

BREAKING 🐞

• The device argument from AttackModel's constructors (i.e., __init__() and classmethod from_pretrained()) have been replaced with the familiar .to(device) method. This also allows for the model to be lazy-loaded to and from GPU devices. Please update accordingly.

New attack register method ✨

• With the newly added @register_attack() decorator, we can now easily register custom attacks that reside outside the standard torchattack library at runtime, and reference the create_attack() method to create custom attacks by their names with the same API.

What's Changed

• Attack registration with decorators, and explicit AttackModel device manipulation API by @spencerwooo in #34

Full Changelog: v1.3.0...v1.4.0

v1.3.0

What's Changed

• Add BSR attack implementation by @spencerwooo in #31
• Add L2T attack implementation by @spencerwooo in #32
• Support for PyTorch 2.6 and Python 3.13 by @spencerwooo in #33

Full Changelog: v1.2.1...v1.3.0

v1.2.1

What's Changed

• [New Attack] ATT attack by @spencerwooo in #29
• [New Attack] GAMA: Generative Adversarial Multi-Object Scene Attacks by @spencerwooo in #30

Full Changelog: v1.2.0...v1.2.1

v1.2.0

What's Changed

• Docs by @spencerwooo in #27
• Update documentation and refactor attack creation process by @spencerwooo in #28

Potential Breaking Changes

In v1.2.0, the create_attack method accepts extra arguments to be passed onto the attack class as keyword arguments, instead of the original attack_args argument.

Full Changelog: v1.0.4...v1.2.0