AdversarialRobustnessReport<T>
Models & Types · AiDotNet.AdversarialRobustness
The empirical adversarial-robustness audit produced by running a configured attack against the trained model. Reports a robustness curve across perturbation budgets rather than a single number, plus a robustness margin — more informative than the single robust-accuracy figure most tools report.
Constructors
| Constructor | Summary |
|---|---|
AdversarialRobustnessReport() |
Properties
| Property | Summary |
|---|---|
IReadOnlyList<RobustnessCurvePoint<T>> RobustnessCurve { get; set; } |
Accuracy versus perturbation budget, in increasing-epsilon order. |
bool SweepAvailable { get; set; } |
Whether the epsilon sweep ran (the attack honored budget changes) or fell back to a single point. |
double AttackSuccessRate { get; set; } |
Fraction of initially-correct samples the attack flips to incorrect at the configured budget. |
double CleanAccuracy { get; set; } |
The model’s accuracy on clean (un-attacked) inputs. |
double CleanVsRobustGap { get; set; } |
Clean accuracy minus robust accuracy at the configured budget — how much the attack costs. |
double MeanPerturbationToFool { get; set; } |
Mean magnitude of the perturbation the attack needed to fool the model (robustness margin) — larger means a more robust model. |
double RobustAccuracy { get; set; } |
The model’s accuracy under attack at the configured perturbation budget. |
double RobustAccuracyAuc { get; set; } |
Area under the robustness curve (normalized), a single summary of robustness across budgets. |
string AttackName { get; set; } |
The configured attack’s type name. |