Synthetic data generation and evaluation for fairness

Thumbnail Image
Bhanot, Karan
Issue Date
Electronic thesis
Computer science
Research Projects
Organizational Units
Journal Issue
Alternative Title
Artificial Intelligence (AI) models often have unfairness, resulting in biased predictions against certain groups of protected individuals. This thesis addresses two broad objectives for quantifying fairness in AI. As the first objective, we define the novel problem of unfairness at the subgroup-level in the context of privacy-preserving synthetic data, especially healthcare data. This is followed by the introduction of time-series and disparate impact fairness metrics for measuring the resemblance (similarity) between the real and synthetic data. For the second objective, we audit the fairness of Machine Learning models and bias mitigation algorithms by stress-testing them under shifts. The thesis describes two auditing pipelines: (a) Fairness Auditor: Grid-based auditing using Iterative Proportional Fitting and (b) Adversarial Auditor: Adversarially attacking utility and fairness objectives using Multi-Objective Bayesian Optimization. Although healthcare data is abundant, access to it is often restricted by privacy laws, and thus, synthetic data provides a viable alternative. Current research measures fairness in various forms but does not discuss the problems of unfairness in synthetic data. Thus, we address the novel problem of defining and then, quantifying the fairness of synthetic data, considering both temporal and non-temporal datasets. Here, we address two definitions of fairness: (a) Machine Learning (ML) fairness of synthetic data and (b) Representational bias in synthetic data. The results highlight that synthetic data exhibits variable bias properties from the real data, as measured by both group fairness metrics on trained ML models and fairness metrics for subgroup-level resemblance. With Machine Learning models and bias mitigation algorithms being used for real-world tasks where shifts in data are common, their applicability under such shifts without robust testing is unknown. This hinders trust in AI models and raises concerns about their utility and fairness under shifts. Here, we discuss how conditional sampling of synthetic data can be used for robust stress-testing of Machine Learning models and bias mitigation algorithms under shifts to identify fairness vulnerabilities (change in biases). Each method is robustly evaluated using the auditing pipelines and summarized using Fairness Reports and novel metrics. These auditing pipelines are a step towards ensuring that Machine Learning methods can be applied under shifts, building trust in these algorithms before real-world application. The results demonstrate that these methods have variable utility and fairness scores on different datasets and lead to increased biases under certain shifts. The thesis concludes with the contributions and discusses potential future work on how the insights can be used to extend Machine Learning auditing to non-binary protected attributes and other tasks such as regression and clustering.
School of Science
Full Citation
Rensselaer Polytechnic Institute, Troy, NY
Terms of Use
PubMed ID