Appendix D, p. 18 — adversarial hinge objective in closed form
ProvedRobustOdds.SVM.advMarginLoss_eqadversarial-robustnessclassificationgaussian-modelp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
For , let be the expectation of the largest hinge loss over perturbations of radius , with the maximum taken separately for each input and label. Then
Thus the adversary contributes the norm of the weights. Adding the same quadratic regularizer to both sides gives the objective identity used in the proof of Lemma D.5.
Formalization Note Equation (2) puts the maximum inside expectation. The display in the proof of Lemma D.5 prints it outside; that placement does not yield this identity.
Preamble
import Mathlib import Definitions.Def_RobustOdds_SVM_Setting
Formal statement
namespace RobustOdds.SVM
open MeasureTheory
theorem advMarginLoss_eq (d : ℕ) (p η ε : ℝ) (hε : 0 ≤ ε)
(w : Fin (d + 1) → ℝ) :
advMarginLoss d p η ε w =
(1 / 2) * ∫ x, max 0 (1 - ∑ i, w i * x i + ε * ∑ i, |w i|)
∂(RobustOdds.Tradeoff.condLaw d p η 1) +
(1 / 2) * ∫ x, max 0 (1 + ∑ i, w i * x i + ε * ∑ i, |w i|)
∂(RobustOdds.Tradeoff.condLaw d p η (-1)) := by sorry
end RobustOdds.SVM
Source
Tsipras et al., Robustness May Be at Odds with Accuracy, arXiv:1805.12152v5, p. 18, App. D, proof of Lemma D.5, two objective displays
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.