Configuration

This page describes how to configure the robustness module that probes how the model behaves under input perturbations.

Inside the robustness key, you can configure one or more named assessors. See Examples for the config shape.

See Supported libraries for the backend behaviour behind _target_, algorithm, and visualiser compatibility.

Options

Name	Allowed	Default	Description
_target_	"TorchattacksAssessor", "FoolboxAssessor", "ImageCorruptionsAssessor"	null	Hydra target for the assessor class.
algorithm	See Supported libraries	null	Name of the underlying attack algorithm to use. The exact class is resolved by the selected assessor backend.
constructor	dict	{}	Keyword arguments forwarded when constructing the assessor / underlying library object. Torchattacks consumes the perturbation budget (eps, alpha, steps, ...) here, since the adapter does attack_class(model, **constructor) once.
call	dict	{}	Keyword arguments forwarded verbatim to the underlying library at call time. Foolbox consumes the perturbation budget (eps, epsilons) here, since foolbox attacks read it at attack(...) time. Any nested dict with a source key is treated as a runtime data source.
raitap	dict	{}	RAITAP-owned runtime options such as batching, progress display, and sample-name metadata. These keys are not forwarded to the underlying library.
raitap.batch_size	int	None	Batch size for generating adversarial examples. If unset, the assessor processes the full input batch in a single call to the attack library. Set this for memory-bound attacks (Square, CW, ...) on large batches.
raitap.show_progress	bool	True	Whether to show a progress bar across attack batches.
raitap.progress_desc	str	null	Description used by the progress bar.
raitap.sample_names	list[str]	null	Optional per-sample names for downstream visualisers. Injected at runtime from the data pipeline. Runtime sample names take precedence over raitap.sample_names from config. The list length must equal the number of input samples N; a mismatch raises raitap.utils.errors.SampleNamesLengthError at factory entry. Omit sample_names to fall back to auto-derived sample ids from the data loader.
raitap.show_sample_names	bool	False	Default toggle for showing sample names in visualiser titles. Set the assessor-level default here under raitap:. If a specific visualiser needs different behaviour, override it with visualisers[].call.show_sample_names.
raitap.input_metadata	dict	null	Input modality + layout hints. Used by image visualisers to refuse non-image results and by the budget norm to size per-sample distance. Auto-inferred from data.source for the default loaders. This per-assessor metadata is scoped to visualiser/budget semantics; backend input reshape is controlled by data.input_metadata.shape instead — see Configuration.
raitap.ci_method	"wilson", "clopper_pearson"	"wilson"	Binomial CI method used by statistical-sampling assessors (ImageCorruptionsAssessor). Ignored by empirical-attack and formal-verification assessors.
raitap.ci_level	float	0.95	Confidence level for the binomial CI. Ignored by empirical-attack and formal-verification assessors.
visualisers	list[dict]	[ImagePairVisualiser]	Visualiser definitions. Each entry must include at least _target_. Each visualiser can also define its own constructor and call blocks. Visualisers declare which AssessmentKind (empirical_attack / formal_verification / statistical_sampling) they support; the factory rejects mismatches at parse time.

Examples

YAML

robustness:
  pgd:
    _target_: "TorchattacksAssessor"
    algorithm: "PGD"
    constructor:
      eps: 0.03
      alpha: 0.0078
      steps: 10
    visualisers:
      - _target_: "ImagePairVisualiser"
  linf_pgd:
    _target_: "FoolboxAssessor"
    algorithm: "LinfPGD"
    constructor:
      rel_stepsize: 0.025
      steps: 40
    call:
      eps: 0.03
    visualisers:
      - _target_: "ImagePairVisualiser"
      - _target_: "PerturbationHeatmapVisualiser"
  avg:
    _target_: "ImageCorruptionsAssessor"
    algorithm: "gaussian_noise"   # one of the 15 ImageNet-C corruptions
    constructor:
      severity: 3                 # 1..5
    raitap:
      ci_method: "wilson"         # or clopper_pearson
      ci_level: 0.95
    visualisers:
      - _target_: "CorruptionAccuracyVisualiser"

Python

from raitap.robustness import corruption_accuracy, foolbox, image_pair, imagecorruptions, perturbation_heatmap, torchattacks

robustness = {
    "pgd": torchattacks(
        algorithm="PGD",
        constructor={"eps": 0.03, "alpha": 0.0078, "steps": 10},
        visualisers=[image_pair()],
    ),
    "linf_pgd": foolbox(
        algorithm="LinfPGD",
        constructor={"rel_stepsize": 0.025, "steps": 40},
        call={"eps": 0.03},
        visualisers=[image_pair(), perturbation_heatmap()],
    ),
    "avg": imagecorruptions(
        algorithm="gaussian_noise",
        constructor={"severity": 3},
        raitap={"ci_method": "wilson", "ci_level": 0.95},
        visualisers=[corruption_accuracy()],
    ),
}

CLI

uv

uv run raitap +robustness=torchattacks robustness.torchattacks.algorithm=PGD robustness.torchattacks.constructor.eps=0.05

pip

raitap +robustness=torchattacks robustness.torchattacks.algorithm=PGD robustness.torchattacks.constructor.eps=0.05