Balancing Fairness, Privacy, and Accuracy: A Multitask Adversarial Framework for Centralized Data-Driven Systems

Technical Summary: Balancing Fairness, Privacy, and Accuracy via a Multitask Adversarial Framework

Problem Statement
Centralized data-driven systems increasingly influence high-stakes sectors such as healthcare, finance, and criminal justice. While algorithmic fairness and data privacy are individually well-studied, their joint optimization remains under-addressed. Current approaches often treat fairness and privacy as separate, post-hoc constraints or sequential processes. This separation creates a fundamental conflict: techniques to enhance privacy (e.g., adding noise via differential privacy or feature anonymization) often degrade data utility and model accuracy. Conversely, methods to enforce fairness (e.g., re-weighting or threshold adjustments) can reduce overall accuracy or inadvertently compromise privacy by altering sensitive attributes in ways that obscure necessary signals. The paper identifies a critical gap in existing literature: the lack of a unified framework that simultaneously optimizes accuracy, fairness, and privacy without significant performance sacrifice, particularly in centralized settings where raw data is accessible.

Methodology: The PFA Framework
The authors propose a novel Privacy-Fairness-Accuracy (PFA) framework, a multitask adversarial model designed to learn a latent representation that hides sensitive attributes while preserving task-relevant information. The architecture consists of three core components:

1. Generator ($G$): Transforms raw input data $X$ into a modified latent representation $X'$. Its objective is to obscure the sensitive attribute $s$ (e.g., gender, race) while retaining information necessary for the primary prediction task $y$.
2. Predictor ($P$): A classifier trained on the transformed data $X'$ to predict the target label $y$. It incorporates a fairness constraint directly into its cost function to ensure equitable outcomes across demographic groups.
3. Discriminator ($D$): An adversarial classifier trained to predict the sensitive attribute $s$ from the latent representation $X'$. The Generator is trained to "fool" the Discriminator, thereby minimizing the information leakage regarding $s$.

Loss Functions and Optimization
The framework optimizes a unified objective function composed of three competing terms, weighted by coefficients $\alpha$, $\beta$, and $\gamma$:

• Prediction Loss ($L_{prediction}$): Minimizes classification error (Binary Cross-Entropy) to maintain accuracy.
• Fairness Loss ($L_{fairness}$): Minimizes the demographic parity gap, defined as the absolute difference in positive prediction rates between protected ($s=1$) and unprotected ($s=0$) groups.
• Privacy Loss ($L_{fooling}$): Maximizes the confusion of the Discriminator (minimizing its ability to predict $s$), effectively acting as an adversarial privacy constraint.

The total Generator loss is formulated as:
$$L_{Generator} = \alpha \cdot L_{prediction} + \beta \cdot L_{fooling} + \gamma \cdot L_{fairness}$$

To address the challenge of balancing these non-linear, coupled objectives, the authors introduce a Modified Adam Optimizer for the cost coefficients. Instead of manual tuning, the weights $\alpha, \beta, \gamma$ are treated as outer-loop design variables. They are dynamically updated based on the gradient of their associated losses relative to the total loss, steering the model toward Pareto-efficient operating points where no single objective can be improved without degrading the others.

Key Contributions

• Unified Framework: The paper introduces a single-end-to-end adversarial architecture that treats privacy and fairness as primary, co-equal objectives rather than post-processing steps.
• Adaptive Weighting: A novel hyperparameter optimization strategy using a modified Adam scheduler to dynamically balance the trade-offs between accuracy, fairness, and privacy, avoiding the instability of naive weighted-sum approaches.
• Comprehensive Evaluation: The framework is validated across five diverse datasets (Adult, Gender Pay Gap, COMPAS, German Credit, and MEPS) covering income, wages, recidivism, credit risk, and healthcare expenditure.
• Robustness Analysis: The study benchmarks PFA against state-of-the-art baselines, including PFairDP, FairDP, and differential privacy mechanisms (Gaussian and Laplacian noise), demonstrating superior adaptability across varying data distributions and class imbalances.

Experimental Results
Extensive testing reveals that the PFA framework achieves a superior balance of the three objectives compared to existing methods:

• Accuracy: PFA achieved the highest accuracy on four of the five datasets (Adult, Gender Pay Gap, German Credit, MEPS) and competitive results on COMPAS.
• Fairness: The model significantly reduced demographic parity gaps. For instance, on the Adult dataset, it achieved a gap of 0.016 (compared to a baseline bias of 0.1945), and on the Gender Pay Gap dataset, it reduced the gap by a factor of seven compared to competitors.
• Privacy: PFA demonstrated strong privacy protection by reducing the predictability of sensitive attributes by a white-box inference attacker (CatBoost) to near-random levels (e.g., ~50-65% accuracy on sensitive attribute prediction, compared to >80% for some baselines).
• Efficiency: The adaptive hyperparameter tuning reduced the time required to find optimal configurations by 16x to >50x compared to exhaustive grid searches.

Significance and Claims
The paper claims that the PFA framework provides a robust solution for the "three-way trade-off" inherent in ethical AI. By integrating fairness and privacy into the training process via adversarial learning, the model avoids the accuracy degradation typically associated with differential privacy or post-hoc fairness corrections. The authors assert that their method offers a more stable and adaptable approach for centralized systems, capable of maintaining high utility while ensuring equitable treatment and protecting sensitive data. The study highlights that while no single metric can be perfectly optimized simultaneously (due to known impossibility bounds), PFA identifies configurations that approach the ideal region of the privacy-fairness-accuracy space more effectively than current state-of-the-art methods. The framework is presented as a practical tool for deploying responsible AI in regulated sectors, enabling compliance with privacy and fairness standards without sacrificing predictive performance.