Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Verification: Abductive Explanations for Post-AI Assessment of Privacy Leakage

Belona Sonna, Alban Grastien, Claire Benn

TL;DR

This paper tackles privacy leakage in AI decision processes by proposing an abductive auditing framework that produces minimal, sufficient explanations for decisions. The core idea is the Potentially Applicable Explanation ($PAE$) and its leakage‑protected variant ($LPPAE$), enabling detection of when open features reveal protected attributes and when they can shield sensitive literals. The authors formalize individual and model‑level leakage, analyze computational complexity (Σ2P‑complete) and propose efficient, albeit exponential, algorithms based on SMT/ SAT formulas. They validate the approach on the German Credit Dataset, showing that some classifiers leak while others do not, and discuss limitations and avenues for scaling and sanitization. The work offers a principled path to reconcile transparency, interpretability, and privacy in deployed AI systems.

Abstract

Privacy leakage in AI-based decision processes poses significant risks, particularly when sensitive information can be inferred. We propose a formal framework to audit privacy leakage using abductive explanations, which identifies minimal sufficient evidence justifying model decisions and determines whether sensitive information disclosed. Our framework formalizes both individual and system-level leakage, introducing the notion of Potentially Applicable Explanations (PAE) to identify individuals whose outcomes can shield those with sensitive features. This approach provides rigorous privacy guarantees while producing human understandable explanations, a key requirement for auditing tools. Experimental evaluation on the German Credit Dataset illustrates how the importance of sensitive literal in the model decision process affects privacy leakage. Despite computational challenges and simplifying assumptions, our results demonstrate that abductive reasoning enables interpretable privacy auditing, offering a practical pathway to reconcile transparency, model interpretability, and privacy preserving in AI decision-making.

Beyond Verification: Abductive Explanations for Post-AI Assessment of Privacy Leakage

TL;DR

This paper tackles privacy leakage in AI decision processes by proposing an abductive auditing framework that produces minimal, sufficient explanations for decisions. The core idea is the Potentially Applicable Explanation () and its leakage‑protected variant (), enabling detection of when open features reveal protected attributes and when they can shield sensitive literals. The authors formalize individual and model‑level leakage, analyze computational complexity (Σ2P‑complete) and propose efficient, albeit exponential, algorithms based on SMT/ SAT formulas. They validate the approach on the German Credit Dataset, showing that some classifiers leak while others do not, and discuss limitations and avenues for scaling and sanitization. The work offers a principled path to reconcile transparency, interpretability, and privacy in deployed AI systems.

Abstract

Privacy leakage in AI-based decision processes poses significant risks, particularly when sensitive information can be inferred. We propose a formal framework to audit privacy leakage using abductive explanations, which identifies minimal sufficient evidence justifying model decisions and determines whether sensitive information disclosed. Our framework formalizes both individual and system-level leakage, introducing the notion of Potentially Applicable Explanations (PAE) to identify individuals whose outcomes can shield those with sensitive features. This approach provides rigorous privacy guarantees while producing human understandable explanations, a key requirement for auditing tools. Experimental evaluation on the German Credit Dataset illustrates how the importance of sensitive literal in the model decision process affects privacy leakage. Despite computational challenges and simplifying assumptions, our results demonstrate that abductive reasoning enables interpretable privacy auditing, offering a practical pathway to reconcile transparency, model interpretability, and privacy preserving in AI decision-making.

Paper Structure

This paper contains 34 sections, 7 theorems, 8 equations, 2 figures, 2 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Preliminaries and Description of the Framework
  4. What Information could be safely disclosed: Data Privacy Dilemma
  5. Notation
  6. Access Levels and Framework Assumptions
  7. Observer Role.
  8. Auditor Role.
  9. Abductive Explanations in Privacy Preserving Decision Systems
  10. Privacy Leakage audit at the individual level
  11. Running Example
  12. Formal Definition of Individual Privacy Leakage
  13. Connection between Privacy Leakage and Abductive Explanation
  14. Theoretical Foundations for Individual Leakage Auditing
  15. Proof:
  16. ...and 19 more sections

Key Result

Theorem 1

There is a $PAE$ for $x$ that does not include the sensitive literal iff there is no leakage for $x$.

Figures (2)

  • Figure 1: The Data Privacy Dilemma: balancing utility, transparency, and privacy in AI-based decision processes
  • Figure 2: Definition of privacy leakage

Theorems & Definitions (18)

  • Definition 1
  • Definition 2: Fully Open Decision
  • Example 1
  • Definition 3: Individual privacy preservation
  • Definition 4: Potentially Applicable Explanation (PAE)
  • Theorem 1
  • Corollary 1
  • Definition 5: Leakage-Protected PAE
  • Example 2
  • Definition 6: Model leakage
  • ...and 8 more