AutoAttackOptions<T>

Options & Configuration · AiDotNet.Models.Options

Configuration for the AutoAttack ensemble (Croce and Hein 2020).

Inheritance: ModelOptions → AdversarialAttackOptions<T> → AutoAttackOptions<T>

For Beginners

You normally only set Epsilon and NormType; the remaining values are the ones the paper used.

How It Works

AutoAttack is parameter-free apart from the threat model: the step size, momentum and step-size schedule of APGD are fixed by the paper. These options select the ensemble version and the per-attack budgets of the reference implementation. Iterations is the iteration budget of APGD and FAB (100 in the paper).

Reference: Croce and Hein (2020), “Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks”, ICML 2020 (https://arxiv.org/abs/2003.01690); reference code fra31/auto-attack.

Constructors

Constructor Summary
AutoAttackOptions() Initializes the options with the paper’s defaults.
AutoAttackOptions(AutoAttackOptions<T> other) Initializes the options by copying every setting of another instance, including the inherited attack settings.

Properties

Property Summary
AutoAttackVersion Version { get; set; } Gets or sets which attack ensemble runs.
double FabAlphaMax { get; set; } Gets or sets FAB’s maximum step bias towards the original point.
double FabBeta { get; set; } Gets or sets FAB’s backward step factor after a successful step.
double FabEta { get; set; } Gets or sets FAB’s overshoot factor of the projection step.
double SquareInitialFraction { get; set; } Gets or sets the initial fraction of the input that one Square attack window covers.
int SquareQueries { get; set; } Gets or sets the query budget of the Square attack.
int TargetClasses { get; set; } Gets or sets how many of the highest-scoring wrong classes the targeted attacks try in turn.