Skip to main content

fgsm

Function fgsm 

Source
pub fn fgsm(
    model: &dyn AttackModel,
    loss: &dyn AttackLoss,
    input: &[f64],
    labels: &[f64],
    config: &AttackConfig,
) -> Result<AdversarialExample, AdversarialError>
Expand description

Fast Gradient Sign Method (Goodfellow et al., 2014).

Computes a single-step adversarial perturbation:

  • L∞: δ = ε · sign(∇_x L)
  • L2: δ = ε · ∇_x L / ‖∇_x L‖₂
  • L1: δ = ε · e_k where k = argmax |∂L/∂xᵢ|