App. C, pp. 15–16 — accuracy against the shift adversary = ½(p(1 + p₊₋ − p₋₊) + (1 − p)(1 + p₋₋ − p₊₊))
ProvedRobustOdds.Tradeoff.shiftAcc_eqadversarial-robustnessclassificationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
In the data model (3) with , let be a measurable classifier, as in the Setting file, and the input after the adversary replaces by for each . Then
Since the adversary turns into and vice versa, the formula is that of the standard accuracy with the second index of every flipped. It is the adversarial half of the reduction of Theorem 2.1 to an inequality in the .
Formalization Note As for the standard-accuracy formula: measurable and -valued, .
Preamble
import Mathlib import Definitions.Def_RobustOdds_Tradeoff_Setting
Formal statement
namespace RobustOdds.Tradeoff
open MeasureTheory ProbabilityTheory
open scoped ENNReal NNReal
/-- App. C, pp. 15–16: the accuracy of a `{±1}`-valued classifier against the adversary that replaces
`G₊` with `G₋` (and vice-versa), expressed in the `p_{ij}`. -/
theorem shiftAcc_eq (d : ℕ) (p η : ℝ) (hp : 1 / 2 ≤ p) (hp1' : p ≤ 1)
(f : (Fin (d + 1) → ℝ) → ℝ) (hf : Measurable f) (hpm : ∀ x, f x = 1 ∨ f x = -1) :
shiftAcc d p η f = 1 / 2 * (p * (1 + pij d η f 1 (-1) - pij d η f (-1) 1) +
(1 - p) * (1 + pij d η f (-1) (-1) - pij d η f 1 1)) := by sorry
end RobustOdds.Tradeoff
Source
Tsipras et al., Robustness May Be at Odds with Accuracy, arXiv:1805.12152v5, pp. 15–16, App. C, proof of Theorem 2.1, display for Pr(f(x_adv) = y)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.