Can Synthetic Data be Fair and Private? A Comparative Study of Synthetic Data Generation and Fairness Algorithms

Technical Summary: Can Synthetic Data be Fair and Private? A Comparative Study of Synthetic Data Generation and Fairness Algorithms

Problem Statement
The increasing integration of Machine Learning (ML) in Learning Analytics (LA) has necessitated robust solutions for algorithmic fairness and data privacy. While Synthetic Data Generators (SDGs) have emerged as dual-purpose tools to enhance privacy (by replacing real data with synthetic counterparts) and improve fairness (by balancing minority group representation), existing literature suggests an inverse relationship between these two objectives. Prior research has largely focused on the trade-offs between privacy and utility or fairness and utility, leaving a gap in understanding the triangular relationship between privacy, fairness, and utility across a broad spectrum of SDGs. Furthermore, while pre-processing fairness algorithms are well-established for real-world data, their efficacy when applied to synthetic data remains underexplored. This study addresses these gaps by investigating which SDGs best balance privacy and fairness and whether pre-processing fairness algorithms can further enhance the fairness of synthetic datasets.

Methodology
The study employs a comparative experimental design utilizing three real-world educational datasets (Student Math performance, Open University Learning Analytics, and Law School Admission Council data) and four sensitive attributes (sex, disability, race).

1. Synthetic Data Generation: Five distinct SDGs were evaluated:

   • CTGAN: A Conditional Tabular GAN focused on utility.
   • DistilGPT2 (DGPT): A Large Language Model adapted for tabular data generation under the GReaT framework.
   • ADSGAN: A GAN-based method using regularization to reduce overfitting and membership inference risks.
   • PATE-GAN: A GAN combined with the Private Aggregation of Teacher Ensembles framework to provide differential privacy guarantees.
   • DECAF: A GAN-based model specifically designed to generate fair synthetic data using causal inference techniques to penalize biased outcomes.

2. Evaluation Metrics:

   • Privacy: Assessed via four metrics: Average Jensen-Shannon Distance (JSD), Wasserstein Distance (WD), Membership Inference Attack (MIA) accuracy (using BNAF), and k-anonymity scores.
   • Fairness: Evaluated using three metrics: Absolute Between Receiver Operating Characteristic curve Area (ABROCA), Error Rate Difference (ERD), and True Positive Rate Difference (TPRD). These were aggregated into a composite fairness score.
   • Utility: Measured by the predictive accuracy (AUC-ROC) of downstream ML models (Random Forest, Logistic Regression, Gaussian Naive Bayes, XGBoost).

3. Experimental Flow:

   • RQ1 (Balance): Synthetic datasets were generated and evaluated for privacy and fairness. Models were trained on these datasets and tested on real data (Same Train, Real Test paradigm) to ensure objective comparison.
   • RQ2 (Improvement): Four pre-processing fairness algorithms (Suppression, Correlation Remover, Disparate Impact Remover, and Reweighing) were applied to the synthetic training data. The resulting models were re-evaluated to measure fairness improvements compared to baseline models and models trained on real data with the same pre-processing.

Key Results

• Privacy-Fairness Trade-off: The study confirms a complex trade-off where optimizing for one often compromises the other. DECAF emerged as the algorithm achieving the best balance between privacy and fairness across all datasets. However, this balance came at a significant cost to utility; DECAF exhibited the lowest predictive accuracy (AUC-ROC) among all generators.
• Algorithmic Specialization:
  • DGPT and CTGAN (utility-focused) showed varying degrees of imbalance, with DGPT leaning slightly more toward fairness and CTGAN showing extreme disparity between privacy and fairness scores.
  • PATE-GAN and ADSGAN (privacy-focused) prioritized privacy, often at the expense of fairness. PATE-GAN, in particular, showed the most extreme trade-offs, achieving high privacy but the lowest fairness scores in some instances.
• Pre-processing Efficacy: Applying pre-processing fairness algorithms to synthetic data yielded greater improvements in fairness than applying them to real data. For instance, the combination of CTGAN with Reweighing (RW) or Suppression (SUP) resulted in significant fairness gains (up to 21.5% improvement on Dataset C).
• Algorithm Sensitivity: The effectiveness of pre-processing algorithms was not uniform; specific fairness techniques performed better with specific SDGs (e.g., CTGAN responded well to SUP and RW, while ADSGAN showed negative results with Correlation Remover on Dataset A).

Significance and Contributions
The paper makes three primary contributions to the field of Learning Analytics and AI ethics:

1. Triangular Relationship Analysis: It provides an extensive empirical study highlighting the triangular relationship between privacy, fairness, and utility. The findings suggest that while privacy and fairness can be jointly optimized (as seen with DECAF), both often exhibit an inverse relationship with utility (predictive accuracy).
2. Superiority of Synthetic + Pre-processing: The study demonstrates that combining synthetic data generation with pre-processing fairness algorithms results in fairer predictions than applying fairness algorithms to real-world data alone. This challenges the notion that synthetic data merely inherits real-world biases without offering a unique pathway for bias mitigation.
3. Practical Guidelines: The research offers actionable recommendations for stakeholders. It suggests that DECAF is optimal for balancing privacy and fairness when utility is secondary, ADSGAN is suitable for privacy-critical scenarios, and DGPT is preferable when fairness is the primary goal. Furthermore, it advocates for a hybrid approach—generating synthetic data followed by pre-processing fairness interventions—as a robust strategy for creating fairer LA models.

Limitations
The authors acknowledge limitations, including the computational constraints that prevented multiple iterations of synthetic data generation (which could have reduced randomness), the focus on a select group of representative SDGs rather than an exhaustive list, and the restriction to tabular data, excluding other LA data types like time series or images.