FairSquare: Probabilistic Verification of Program Fairness Aws - PowerPoint PPT Presentation

FairSquare: Probabilistic Verification of Program Fairness Aws Albarghouthi Loris D’Antoni Samuel Drews University of Wisconsin-Madison Aditya V. Nori Microsoft Research

Machine Data Model Learning

Machine Data Model Learning small error

Machine Data Model Learning small error racist? sexist?

Group Fairness decision- making program

Group Fairness sensitive feature (e.g. minority)

Group Fairness sensitive feature (e.g. minority)

Group Fairness population model

Individual Fairness similarity

Q: What are good definitions of fairness?

Q: What are good definitions of fairness? A: Someone else’s problem

Contributions ● Formalize fairness definitions as probabilistic verification problems ● Decision procedure for evaluating probabilistic postconditions

FairSquare

Population Model Decision Program Fairness FairSquare Definition

Population Model Unfairness proof Decision Program Fairness proof Fairness FairSquare Definition

population decision-making model program

population decision-making model program

represent assignments as a region (LRA formula) colRank “weighted volume” ethnicity yExp

define dec(colRank, yExp): define popModel(): expRank ← yExp – colRank ethnicity ~ gauss(0,10) if colRank <= 5: colRank ~ gauss(25,10) hire ← true yExp ~ gauss(10,5) elif expRank > -5: if ethnicity > 10: hire ← true colRank ← colRank + 5 else: return colRank, yExp hire ← false return hire popModel dec Pr

define dec(colRank, yExp): define popModel(): expRank ← yExp – colRank ethnicity ~ gauss(0,10) if colRank <= 5: colRank ~ gauss(25,10) hire ← true yExp ~ gauss(10,5) elif expRank > -5: if ethnicity > 10: hire ← true colRank ← colRank + 5 else: return colRank, yExp hire ← false return hire popModel dec Pr

define dec(colRank, yExp): define popModel(): expRank ← yExp – colRank ethnicity ~ gauss(0,10) if colRank <= 5: colRank ~ gauss(25,10) hire ← true yExp ~ gauss(10,5) elif expRank > -5: if ethnicity > 10: hire ← true colRank ← colRank + 5 else: return colRank, yExp hire ← false return hire popModel dec Pr

Evaluating post ● Obtain formula for each probability in post ● Underapproximate the weighted volume of ● Overapproximate by doing the same for eventually, approximations are good enough

“Live” demo

Case Study Bayesian Population Networks Model Income Dataset [1] Decision Trees, Decision Linear SVMs, Program RLU Neural Nets Group Fairness Fairness for Women Definition [1] https://archive.ics.uci.edu/ml/datasets/adult

Fairness Verification Problems 14 12 NN 10 SVM 8 DT 6 4 2 0 Total FairSquare PSI [1] VolComp [2] [1] Gehr et al. CAV 2016 [2] Sankaranarayanan et al. PLDI 2013

Population Model Unfairness proof Decision Program Fairness proof Fairness FairSquare Definition