AIRT Image Attack Cleanup

Key Changes:

• Add 4 new image attack implementations (HopSkipJump, NES, SimBA, ZOO)
• Refactor attack/search separation: move search logic to airt/search/
• Add dreadnode attack CLI for running pre-configured attacks
• Add dreadnode task CLI for task introspection and execution
• Implement ZOO hierarchical attack with coordinate sampling
• Add NES gradient estimation via random direction probing
• Enhance eval output formatting with score tables and trial views
• Add importance sampling and ADAM optimizer support for ZOO
• Improve boundary search in HopSkipJump with binary line search
• Extend meta introspection to support task callable discovery
• Add format.py utilities for console table rendering
• Refactor optimization/sampling.py for reusable coordinate sampling
• Update study formatting to display multi-objective results
• Fix prompt attack to use parent_depth instead of context_depth
• Add codespell exception for "asend" keyword
• Extend documentation for all new attack methods and CLI commands

Generated Summary:

• Updated the .pre-commit-config.yaml to include additional word exclusions in the codespell hook.
• Introduced several new image attack methods in the documentation:
  • HopSkipJump Attack: Adds functionality for black-box image classifiers with detailed parameter descriptions.
  • NES Attack: Implements Natural Evolution Strategies with parameters for optimization steps and exploration variance.
  • SimBA Attack: Introduces Simple Black-box Attack with support for random perturbation masks.
  • ZOO Attack: A Zeroth-Order Optimization attack allowing for enhanced black-box image classification.
• Adjusted the prompt_attack function by renaming context_depth to parent_depth for clarity regarding the parameter's purpose.
• Enhanced the hop_skip_jump_search and respective classes to support new optimization parameters and improved logging during search operations.
• Overall improvements may lead to better performance and capabilities in generating adversarial examples for various neural network architectures.

This summary was generated with ❤️ by rigging