Fair Synthetic Data

The foundational property ensuring that the joint distribution of features and labels in the synthetic dataset is independent of the sensitive attribute. This is achieved by enforcing constraints that equalize the probability of a favorable outcome across all demographic groups.

• Breaks the correlation between protected attributes and labels
• Uses Wasserstein distance minimization to match group-specific distributions
• Example: Loan approval rates in synthetic data are identical across all ZIP codes, removing historical redlining patterns

Fair synthetic data generators inject calibrated Laplacian or Gaussian noise into the training procedure, providing a mathematical guarantee that the presence or absence of any single individual's record does not significantly alter the generated output.

• Bounded by a privacy loss parameter epsilon (ε)
• Prevents membership inference attacks against the original data
• Example: A synthetic medical dataset where ε=1.0 ensures an adversary cannot determine if a specific patient was in the original study

Unlike naive oversampling, fair synthetic data retains the true causal graph of the domain while severing non-causal paths from sensitive attributes. This ensures the model learns legitimate decision rules rather than spurious correlations.

• Uses causal discovery algorithms to map feature relationships
• Removes confounding effects of protected attributes
• Example: Synthetic hiring data preserves the causal link between 'years of experience' and 'salary' but removes the spurious link between 'gender' and 'salary'

A multi-objective optimization process that maximizes the synthetic data's usefulness for downstream model training while minimizing the discrimination score. This is quantified through the fairness-utility Pareto frontier.

• Measures fidelity via propensity score matching
• Evaluates utility through Train on Synthetic, Test on Real (TSTR) benchmarks
• Example: A credit scoring model trained on fair synthetic data achieves 95% of the original AUC while reducing disparate impact by 100%

Modern fair synthesizers employ a min-max game between a generator and a critic. The generator creates realistic records, while an adversarial critic penalizes the generator if it can predict the sensitive attribute from the synthetic data.

• Based on the FairGAN architecture
• The generator's loss function includes a fairness penalty term
• Example: A generator creates synthetic customer profiles; the adversary fails to predict race from purchase history, confirming successful obfuscation

Ensures that the synthetic data provides sufficient statistical representation for all intersectional subgroups, not just the majority. This prevents the erasure of minority patterns under the guise of fairness.

• Validates minimum sample counts for intersectional cohorts
• Prevents fairness gerrymandering where subgroups are ignored
• Example: A synthetic dataset for a global retailer contains robust transaction patterns for 'women over 60 in rural areas,' a segment often lost in aggregation

The systematic process of applying algorithmic techniques to reduce unwanted systematic errors in machine learning models. When generating fair synthetic data, mitigation is applied during the generation process itself.

• Pre-processing: Modifying the original dataset before synthesis to remove historical biases.
• In-processing: Embedding fairness constraints directly into the generative model's loss function.
• Post-processing: Adjusting the generated synthetic dataset to meet parity targets.

A group fairness metric requiring a model's positive prediction rate to be equal across all demographic groups. In the context of fair synthetic data, this constraint is explicitly enforced during generation.

• Ensures statistical independence from the sensitive attribute.
• The synthetic dataset is engineered so that P(ŷ=1|A=a) = P(ŷ=1|A=b) for all groups a and b.
• Often used as a hard constraint in fair representation learning models like adversarial autoencoders.

A causal definition of fairness where a decision is considered fair if it would remain the same in a counterfactual world where an individual's sensitive attributes were different. Fair synthetic data generators use this principle to create plausible alternate realities.

• Relies on Structural Causal Models (SCMs) to model the data-generating process.
• Enables counterfactual data augmentation: generating synthetic twins where only the protected attribute is altered.
• Provides stronger fairness guarantees than purely associational methods.

A pre-processing approach that learns a latent data representation encoding useful information for prediction while obfuscating information about sensitive attributes. This is a core mechanism for generating fair synthetic data.

• Uses adversarial networks to train an encoder that maximizes target prediction accuracy while a discriminator fails to predict the protected attribute.
• The resulting latent space can be sampled to generate new, bias-free records.
• Techniques include Variational Fair Autoencoders (VFAE) and adversarial debiasing.

A legal and quantitative measure of discrimination occurring when a facially neutral policy disproportionately harms members of a protected group. Fair synthetic data is explicitly designed to eliminate disparate impact from downstream models.

• Assessed using the 80% rule: the selection rate for a protected group must be at least 80% of the rate for the group with the highest selection rate.
• Synthetic data generators can be programmed to guarantee a disparate impact ratio of 1.0.
• Critical for compliance with lending, hiring, and credit regulations.

The inherent tension in model optimization where enforcing strict fairness constraints often results in a measurable reduction in overall predictive accuracy. Fair synthetic data aims to minimize this trade-off.

• By correcting bias at the data level, downstream models can often achieve higher utility for the same fairness level.
• The trade-off is quantified using Pareto frontiers plotting accuracy against metrics like equalized odds.
• Advanced generative models like fair GANs and fair diffusion models push the frontier closer to the ideal of no trade-off.