App. C, p. 15 — standard accuracy = ½(p(1 + p₊₊ − p₋₋) + (1 − p)(1 + p₋₊ − p₊₋))
ProvedRobustOdds.Tradeoff.stdAcc_eqadversarial-robustnessclassificationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
In the data model (3) with , let be a measurable classifier, and let () be the probability that predicts when (with read as ) and the weak features follow . Then the standard accuracy of is
Conditioning on and on splits the accuracy into four cases, each a or its complement. Together with the companion formula for the shift adversary, this turns Theorem 2.1 into an inequality between two linear expressions in the .
Formalization Note Measurability of and for every are the paper's "classifier that maps an input to a class in "; they make . The upper bound is part of being a probability.
Preamble
import Mathlib import Definitions.Def_RobustOdds_Tradeoff_Setting
Formal statement
namespace RobustOdds.Tradeoff
open MeasureTheory ProbabilityTheory
open scoped ENNReal NNReal
/-- App. C, p. 15: the standard accuracy of a `{±1}`-valued classifier expressed in the `p_{ij}`. -/
theorem stdAcc_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) :
stdAcc 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, p. 15, App. C, proof of Theorem 2.1, display for Pr(f(x) = y)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.