App. C, p. 14 — subtracting 2ηy from each weak feature turns G_y into G_{−y}
ProvedRobustOdds.Tradeoff.weakLaw_shiftadversarial-robustnessgaussian-modelp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , , and for let be the law of the weak features in the data model (3). Let , , subtract from every coordinate. Then for every
For this says the adversary with changes to , and for it changes to . This is the mechanism behind Theorem 2.1: a perturbation of size makes the weak features look as if they came from the other class.
Preamble
import Mathlib import Definitions.Def_RobustOdds_Tradeoff_Setting
Formal statement
namespace RobustOdds.Tradeoff
open MeasureTheory ProbabilityTheory
open scoped ENNReal NNReal
/-- App. C, p. 14: the adversary that replaces `x_i` by `x_i − 2ηy` (`i ≥ 2`) maps the law of the
weak features under label `y` to their law under label `−y`; for `y = 1` it changes `G₊` to `G₋`,
for `y = −1` it changes `G₋` to `G₊`. -/
theorem weakLaw_shift (d : ℕ) (η : ℝ) :
∀ y : ℝ, (weakLaw d η y).map (fun z i => z i - 2 * η * y) = weakLaw d η (-y) := by sorry
end RobustOdds.Tradeoff
Source
Tsipras et al., Robustness May Be at Odds with Accuracy, arXiv:1805.12152v5, p. 14, App. C, proof of Theorem 2.1, first paragraph
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.