Fair Aggregation (q-FFL)

The q hyperparameter tunes the aggressiveness of fairness enforcement. When q=0, q-FFL reduces to standard FedAvg, optimizing for average performance. As q increases, the objective function penalizes high-loss clients more heavily, forcing the global model to prioritize accuracy uniformity over raw aggregate performance. This creates a tunable trade-off between overall model accuracy and inter-client performance variance.

q-FFL modifies the standard federated optimization objective by raising each client's local loss to the power of (q+1) before summation. This mathematical transformation amplifies the gradient contribution from clients experiencing higher empirical risk. The server solves: min f_q(w) = Σ (F_k(w))^(q+1) / (q+1), where F_k is the local loss of client k. This ensures underperforming nodes exert proportionally greater influence on the global update direction.

Unlike FedAvg which weights updates solely by dataset size, q-FFL implicitly allocates more optimization resources to statistically challenging or underrepresented client populations. This is critical in healthcare settings where rare disease cohorts or smaller clinics with limited data would otherwise be ignored by a model optimizing for aggregate accuracy. The mechanism prevents representation bias without requiring explicit demographic labels.

q-FFL employs a gradient-based optimization approach that guarantees convergence to a Pareto-stationary point. The algorithm uses Lipschitz smoothness assumptions to bound the gradient norm. During each communication round, the server computes a scaled global gradient that balances fairness with convergence speed. Empirical results on federated benchmarks like FEMNIST and CelebA demonstrate that q-FFL reduces accuracy variance across clients by up to 40% compared to FedAvg, with minimal degradation to mean accuracy.

By penalizing high-loss clients, q-FFL indirectly mitigates client drift in heterogeneous environments. When local data distributions are non-IID, standard FedAvg allows clients with easy distributions to dominate the global model. q-FFL counteracts this by re-weighting the aggregation step, ensuring that clients with difficult or skewed data distributions—common in clinical settings with varying patient demographics—maintain influence throughout training. This complements techniques like FedProx and SCAFFOLD.

The primary output of q-FFL is a global model that exhibits uniform accuracy distribution across participating clients at inference time. This is measured using metrics like worst-10% accuracy and coefficient of variation of client accuracies. In healthcare federated learning, this translates to diagnostic models that perform consistently whether deployed at a large urban hospital or a rural clinic, addressing health equity concerns without requiring post-hoc calibration or separate per-site models.

A robust aggregation framework that optimizes the global model for the worst-case mixture of client data distributions. Unlike q-FFL which re-weights the objective to reduce variance, AFL provides minimax performance guarantees even under adversarial data shifts.

• Targets the maximum expected loss across any client mixture
• Provides theoretical robustness against distributional drift
• Complements q-FFL's fairness-by-reweighting with fairness-by-optimization

An aggregation framework that partitions clients into distinct clusters based on data distribution similarity, maintaining separate global models for each cluster. While q-FFL enforces uniform performance across all clients, IFCA acknowledges irreducible heterogeneity by grouping similar clinical populations.

• Uses iterative clustering to identify natural data groupings
• Maintains K independent global models for K clusters
• Alternative to q-FFL when client populations are fundamentally distinct

Uses Moreau envelopes to decouple personalized model optimization from global model learning. Unlike q-FFL which enforces a single fair global model, pFedMe allows each client to maintain a local model that is regularized toward the global consensus.

• Balances global fairness with local adaptation
• L2 regularization between local and global parameters
• Addresses the fairness-personalization tradeoff in clinical settings

A dynamic regularization method that aligns local and global stationary points by adjusting the local loss function with a penalty term based on the historical gradient of the global model. FedDyn and q-FFL both modify the local objective, but with different goals.

• FedDyn targets convergence consistency across heterogeneous clients
• q-FFL targets performance distribution uniformity
• Both techniques can be combined for convergent and fair federated optimization

The standard federated strategy where local model updates are combined proportionally to dataset size. This is the baseline that q-FFL explicitly improves upon, as naive weighting can bias the global model toward data-rich institutions.

• Update weight = n_k / Σ n_i where n_k is local dataset size
• Large hospitals dominate; rare disease clinics are marginalized
• q-FFL's re-weighting directly counters this majority-class bias

Views each client's learning problem as a distinct but related task, leveraging multi-task learning principles to share statistical strength. While q-FFL enforces a single fair model, FedMT embraces task diversity.

• Learns shared feature representations across clients
• Allows task-specific output layers for each institution
• Suitable when clinical endpoints differ fundamentally across sites