Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causally Perturbed Fairness Testing

Chengwen Du, Tao Chen

TL;DR

CausalFT addresses the challenge of uncovering fairness bugs in AI models handling tabular data by introducing a causal perturbation framework. It builds a direct causal graph using LiNGAM from training data, identifies the most causally relevant non-sensitive feature f_c to a sensitive feature f_s, and uses the do-calculus to quantify their joint influence on the label. This causality-guided relationship is injected into perturbation-based generators, relaxing the definition of discriminatory instances and allowing invalid pairs to be repaired, resulting in up to 93% improvement in bug discovery across 1296 trials with acceptable overhead. Compared with correlation-based baselines, CausalFT demonstrates superior effectiveness and generalizes across multiple models and datasets, enabling more robust bias detection and potential improvements after retraining with generated samples.

Abstract

To mitigate unfair and unethical discrimination over sensitive features (e.g., gender, age, or race), fairness testing plays an integral role in engineering systems that leverage AI models to handle tabular data. A key challenge therein is how to effectively reveal fairness bugs under an intractable sample size using perturbation. Much current work has been focusing on designing the test sample generators, ignoring the valuable knowledge about data characteristics that can help guide the perturbation and hence limiting their full potential. In this paper, we seek to bridge such a gap by proposing a generic framework of causally perturbed fairness testing, dubbed CausalFT. Through causal inference, the key idea of CausalFT is to extract the most directly and causally relevant non-sensitive feature to its sensitive counterpart, which can jointly influence the prediction of the label. Such a causal relationship is then seamlessly injected into the perturbation to guide a test sample generator. Unlike existing generator-level work, CausalFT serves as a higher-level framework that can be paired with diverse base generators. Extensive experiments on 1296 cases confirm that CausalFT can considerably improve arbitrary base generators in revealing fairness bugs over 93% of the cases with acceptable extra runtime overhead. Compared with a state-of-the-art approach that ranks the non-sensitive features solely based on correlation, CausalFT performs significantly better on 64% cases while being much more efficient. Further, CausalFT can better improve bias resilience in nearly all cases.

Causally Perturbed Fairness Testing

TL;DR

CausalFT addresses the challenge of uncovering fairness bugs in AI models handling tabular data by introducing a causal perturbation framework. It builds a direct causal graph using LiNGAM from training data, identifies the most causally relevant non-sensitive feature f_c to a sensitive feature f_s, and uses the do-calculus to quantify their joint influence on the label. This causality-guided relationship is injected into perturbation-based generators, relaxing the definition of discriminatory instances and allowing invalid pairs to be repaired, resulting in up to 93% improvement in bug discovery across 1296 trials with acceptable overhead. Compared with correlation-based baselines, CausalFT demonstrates superior effectiveness and generalizes across multiple models and datasets, enabling more robust bias detection and potential improvements after retraining with generated samples.

Abstract

To mitigate unfair and unethical discrimination over sensitive features (e.g., gender, age, or race), fairness testing plays an integral role in engineering systems that leverage AI models to handle tabular data. A key challenge therein is how to effectively reveal fairness bugs under an intractable sample size using perturbation. Much current work has been focusing on designing the test sample generators, ignoring the valuable knowledge about data characteristics that can help guide the perturbation and hence limiting their full potential. In this paper, we seek to bridge such a gap by proposing a generic framework of causally perturbed fairness testing, dubbed CausalFT. Through causal inference, the key idea of CausalFT is to extract the most directly and causally relevant non-sensitive feature to its sensitive counterpart, which can jointly influence the prediction of the label. Such a causal relationship is then seamlessly injected into the perturbation to guide a test sample generator. Unlike existing generator-level work, CausalFT serves as a higher-level framework that can be paired with diverse base generators. Extensive experiments on 1296 cases confirm that CausalFT can considerably improve arbitrary base generators in revealing fairness bugs over 93% of the cases with acceptable extra runtime overhead. Compared with a state-of-the-art approach that ranks the non-sensitive features solely based on correlation, CausalFT performs significantly better on 64% cases while being much more efficient. Further, CausalFT can better improve bias resilience in nearly all cases.
Paper Structure (52 sections, 13 equations, 6 figures, 14 tables, 1 algorithm)

This paper contains 52 sections, 13 equations, 6 figures, 14 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Model Fairness
  4. Fairness Testing and Problem Formulation
  5. Observations and Motivation
  6. Fairness Testing with Causal Perturbation
  7. Feature Causality Analysis
  8. Causal Graph Extraction
  9. Causal Effects Computation
  10. Causal Perturbation
  11. Common Parts of Perturbation
  12. Seamless Causality Injection
  13. Invalidity Repair
  14. Experiments
  15. Datasets
  16. ...and 37 more sections

Figures (6)

  • Figure 1: Exampled partial dependence plots for the most significant interactions between the sensitive and non-sensitive features in DNN prediction. (a), (b), and (c) are the Adult dataset when Age is the sensitive feature; (d), (e), and (f) are the German dataset when Age is the sensitive feature; (g), (h), and (i) are the Compas dataset when Race is the sensitive feature. The numbers in the figures denote the inferred probability belongs to the positive class.
  • Figure 2: Workflow of CausalFT for fairness testing.
  • Figure 3: The causal graph built for the Adult dataset (Age being the only concerned sensitive feature).
  • Figure 4: Comparison of the Scott-Knott test ranks when pairing CausalFT with different generators on all models and cases.
  • Figure 5: Comparing runtime efficiency.NeuronFair (b) and SG (c) are the fastest and slowest generators for DNN5; NeuronFair (d) and SG (e) are the fastest and slowest generators for DNN6; NeuronFair (f) and ExpGA (g) are the fastest and slowest generators for LR; EIDIG (h) and SG (i) are the fastest and slowest generators for RF, respectively (mean/deviation over all possible sensitive features/runs).
  • ...and 1 more figures