Federated Learning with Fairness Constraints

These are modifications to the core federated averaging (FedAvg) algorithm that explicitly account for client performance disparities during the server-side model update. Instead of a simple weighted average, the aggregation function incorporates fairness metrics.

• q-Fair Federated Learning (q-FFL): Assigns higher weight to clients with larger local loss values, prioritizing the improvement of the worst-performing clients.
• Agnostic Federated Learning (AFL): Optimizes for a distributionally robust objective, seeking a model that performs well over any possible mixture of client data distributions.
• TERM (Tilted Empirical Risk Minimization): Applies a tilt to the loss function, amplifying the influence of higher-loss clients to steer the global model toward more equitable solutions.

Quantifiable measures are essential to diagnose and constrain unfairness. These metrics evaluate the spread of model performance across the client population.

• Worst-Case Client Accuracy: The minimum accuracy achieved by any client on their local test set. Directly targets the most disadvantaged participant.
• Variance of Client Accuracies: The statistical variance of performance metrics (accuracy, F1-score) across all clients. A lower variance indicates more uniform performance.
• Jain's Fairness Index: A bounded metric (between 0 and 1) that measures the equality of a resource distribution; here, it assesses how evenly performance is distributed, where 1 indicates perfect fairness.
• Difference from Mean Performance: The absolute gap between individual client performance and the global average, highlighting outliers.

This approach adds explicit penalty terms to the local or global training objective to discourage models from overfitting to dominant data distributions and neglecting minority clients.

• Local Fairness Regularizers: Clients add a term to their loss function that penalizes predictions which are biased against protected attributes (e.g., demographic groups) present in their local data.
• Global Disparity Regularization: The server adds a regularization term during aggregation that penalizes the global model based on the variance of client losses reported in the round, actively reducing performance spread.
• Gradient Correction for Fairness: Techniques that project or adjust client gradients before aggregation to align them with a direction that reduces performance disparity, often derived from constrained optimization principles.

Algorithms control which clients participate in each training round and how much influence their updates have, based on fairness objectives rather than just data quantity.

• Active Fairness-Aware Sampling: The server probabilistically selects clients for a training round, favoring those with historically lower performance or higher loss to give them more training influence.
• Inverse Loss-Based Weighting: During aggregation, a client's model update is weighted inversely to its reported loss or performance, boosting the contribution of struggling clients.
• Power-of-Choice Selection: A hybrid method that selects a subset of clients with the highest local loss from a larger random pool, balancing exploration with targeted improvement for fairness.

Producing a single, perfectly fair global model for all Non-IID clients is often impossible. Personalization techniques provide an alternative path to equity by tailoring models to individual clients.

• Model Interpolation: Each client receives a model that is a weighted mixture of the global (fairness-constrained) model and a locally fine-tuned model, allowing adaptation without sacrificing collaborative fairness gains.
• Fair Meta-Learning: A meta-learning framework (e.g., using MAML) is used to learn a global model initialization that can be quickly and effectively fine-tuned by any client to achieve high performance, ensuring all clients have a strong starting point.
• Fair Clustered Personalization: Clients are clustered based on data distribution and performance profiles. Separate global models are trained per cluster, ensuring fairness within more homogeneous groups.

Specialized tools are required to simulate and measure fairness in federated environments, as traditional centralized test sets are insufficient.

• Federated Fairness Evaluation: Model performance is evaluated on each client's local hold-out test set. The suite of disparity metrics (worst-case accuracy, variance, etc.) is computed across all clients to produce a fairness report.
• Synthetic Fairness Benchmarks: Datasets like LEAF are partitioned using biased sampling strategies (e.g., correlating labels with synthetic protected attributes) to create controlled Non-IID scenarios with known fairness gaps.
• Fairness vs. Accuracy Trade-off Curves: Critical analysis involves plotting the global model's average accuracy against a fairness metric (e.g., worst-case client accuracy) across different algorithm hyperparameters, visualizing the inherent trade-off.

Collaborative training of fraud detection models among competing financial institutions or e-commerce platforms. Each client's data contains unique fraud patterns specific to their business model and user base. Fairness constraints ensure that the collaborative model provides balanced benefit to all participants, preventing free-riding where some clients benefit more than others contribute.

• Key Challenge: Incentive misalignment; participants may drop out if the global model primarily serves the largest entity's data distribution.
• Solution: Implement contribution-aware fairness, where the training process is constrained to ensure model utility is proportional to each client's data contribution and needs.

Governments or NGOs training models to optimize resource allocation (e.g., healthcare, disaster relief, educational funding) using data from different districts or departments. Data is Non-IID due to socioeconomic disparities across regions. Fairness constraints are legally and ethically mandated to ensure the model does not exacerbate existing inequalities.

• Key Challenge: Achieving equitable performance (e.g., prediction accuracy for service needs) across wealthy and under-resourced districts.
• Methodology: Integrate fairness regularizers (e.g., based on statistical parity difference) into the local client loss function before updates are shared and aggregated.

Agnostic Fairness is a fairness criterion in federated learning where the goal is to ensure the global model performs well for any target distribution formed by a mixture of the clients' local data distributions. This is a stronger guarantee than average accuracy, directly addressing performance disparities caused by Non-IID data.

• Key Mechanism: Often formalized as minimizing the worst-case loss over a convex hull of client distributions.
• Relation to FL: Algorithms like Agnostic Federated Learning (AFL) optimize for this criterion, preventing the model from being biased toward clients with larger or easier datasets.

q-Fair Federated Learning is an optimization framework that aims to improve the performance of the worst-performing quantile (e.g., the bottom 10% or 25%) of clients. Instead of pure worst-case (minimax) optimization, it provides a tunable fairness-utility trade-off.

• Key Mechanism: Optimizes a Conditional Value at Risk (CVaR)-like objective focused on the tail of the client performance distribution.
• Practical Impact: More robust than average accuracy and more flexible than strict minimax, allowing system designers to control how aggressively to prioritize struggling clients.

This approach reformulates the federated learning objective to maximize the Area Under the ROC Curve (AUC), a metric that is inherently more sensitive to class-imbalanced data common in Non-IID settings.

• Why AUC? AUC measures the model's ability to rank positive instances higher than negative ones, which is more stable across different client label distributions than accuracy.
• Federated Challenge: Requires decentralized computation of pairwise loss functions, leading to specialized communication-efficient algorithms for federated AUC maximization.

Fairness-Aware Client Selection algorithms modify the sampling strategy for each training round to prioritize clients that are historically underserved or whose data will most improve fairness metrics.

• Methods Include: Power-of-choice selection biasing towards clients with higher loss, or bandit-based algorithms that explore client selection to reduce performance variance.
• System Trade-off: Balances the goal of fairness against practical constraints like client availability and communication costs.

These are the quantitative measures used to evaluate fairness in a federated model. Key metrics include:

• Worst-Case Client Accuracy: The minimum accuracy across all clients.
• Accuracy Variance: The statistical variance of accuracy across the client population.
• Jain's Fairness Index: A bounded metric (between 0 and 1) that measures the equality of a resource distribution, applied here to model performance.
• Delta Disparity: The difference between the best and worst client performance.

Monitoring these metrics is essential for diagnosing and mitigating unfairness arising from data heterogeneity.

FairFed is a specific algorithm for fair federated learning that employs an adaptive aggregation rule at the server. The server assigns higher weights to client updates from historically lower-performing groups during the model averaging step.

• Core Mechanism: Uses the performance history of client groups (e.g., demographic groups or data clusters) to dynamically adjust their influence on the global model.
• Outcome: Actively reduces performance disparities between groups across training rounds, moving beyond static, uniform averaging (FedAvg).