Bias Mitigation

This method divides the client population into non-overlapping subgroups (strata) based on key attributes—such as geographic region, device type, or inferred data distribution—and samples a proportional number of clients from each stratum for every training round.

• Purpose: Guarantees that all predefined subgroups are represented, preventing the model from being dominated by the most common or readily available client type.
• Example: In a healthcare FL application, strata could be hospitals of different sizes (urban, rural) to ensure the model learns from diverse patient demographics, not just large urban centers.
• Implementation: Requires a profiling phase to categorize clients, followed by a sampling rule (e.g., sample 2 clients from each of 5 strata per round).

This approach modifies the standard utility function used for client scoring by incorporating fairness constraints. Instead of selecting clients solely based on metrics like gradient norm or dataset size, it adds a penalty for repeatedly selecting the same clients or a bonus for underrepresented groups.

• Core Mechanism: The selection score becomes a weighted sum: Score = α * (Statistical Utility) + β * (Fairness Term).
• Fairness Terms: Common terms include:
  • Round Robin Penalty: Reduces the score of clients selected in recent rounds.
  • Demographic Parity Reward: Increases the score for clients from groups with low historical participation rates.
• Outcome: Balances model convergence speed with long-term participation equity, mitigating representation bias.

This is a pre-processing strategy where clients are grouped into clusters based on the similarity of their local data distributions (e.g., using metrics on model updates or metadata). Selection then ensures picks are made from multiple clusters.

• Process:
  1. Perform an initial profiling round to collect client updates or data statistics.
  2. Use clustering algorithms (e.g., K-means on gradient embeddings) to group clients.
  3. For each training round, select a subset of clients that maximizes the number of clusters represented.
• Advantage: Actively promotes client diversity without requiring explicit demographic labels, which may be private or unknown. It mitigates bias stemming from non-IID data by forcing the model to learn from multiple data modalities in each round.

This framework treats client selection as an online learning problem. Each client is an "arm" of a bandit. The algorithm must balance exploitation (selecting clients known to provide high-quality updates) with exploration (selecting less-known clients to discover potentially valuable data).

• Mitigates Bias: Inherently prevents over-reliance on a stable set of "good" clients, which can introduce availability bias. The exploration force systematically samples from the long tail of the client population.
• Algorithms: Upper Confidence Bound (UCB) or Thompson Sampling are commonly adapted. The "reward" can be the reduction in global loss or the magnitude of the client's update.
• Result: Over time, the algorithm builds a more accurate model of all clients' value, leading to selections that are both efficient and less biased.

A federated coreset is a small, weighted synthetic dataset maintained on the server that approximates the global data distribution. It is used to guide client selection by identifying which clients' local data best compensates for gaps in the current global model's knowledge.

• Mechanism:
  1. The server uses the coreset to identify data regions (in feature space) where the model is currently weak.
  2. It then selects clients whose local data statistics (e.g., via secure, privacy-preserving sketches) indicate they can contribute samples from those underrepresented regions.
• Impact: Directly counteracts coverage bias by proactively seeking data that fills distributional holes. It moves selection beyond simple metrics like data quantity to data relevance and complementarity.

Bias can stem from participation bias, where only clients with abundant resources (power, bandwidth) volunteer. Incentive mechanisms use economic or reputational rewards to encourage a broader, more representative set of clients to make themselves available for selection.

• Methods:
  • Performance-based Rewards: Clients receive higher rewards (e.g., tokens, improved model personalization) for contributions that uniquely improve model fairness or accuracy on underrepresented classes.
  • Fairness-aware Payouts: The reward function explicitly compensates clients from rare strata, making it worthwhile for them to participate despite potential resource costs.
• Goal: Aligns individual client rationality with the system goal of a unbiased, robust global model by shaping the pool of eligible clients.

A client selection approach that incorporates explicit fairness constraints to prevent systematic underrepresentation. This is a direct implementation of bias mitigation.

• Key Methods: Enforcing quotas for demographic groups, using fairness regularizers in the utility function, or applying constrained optimization.
• Objective: Ensures devices or data distributions from protected classes (e.g., specific geographic regions, device models) are selected with a guaranteed minimum frequency.
• Example: A policy guaranteeing at least 10% of selected clients per round come from a region with historically poor connectivity.

A statistical sampling method used as a foundational technique for bias mitigation in client selection.

• Process: The client population is divided into non-overlapping subgroups (strata) based on key attributes like data distribution, device type, or geographic location.
• Selection: Clients are then sampled from each stratum, proportionally or with adjusted weights, to ensure the selected cohort is representative of the overall population.
• Contrast with Random: Unlike pure random selection, which can lead to skewed cohorts by chance, stratified sampling provides deterministic guarantees against representation bias.

A core objective and metric for bias mitigation, focusing on the statistical and systemic variety within a selected cohort.

• What it Measures: The degree to which the selected clients' data distributions and characteristics collectively represent the global target population.
• Why it Matters: Low diversity can cause the global model to overfit to dominant patterns, failing to generalize for underrepresented groups.
• Quantification: Often measured using metrics like client distribution divergence (e.g., KL divergence between cohort and global data distributions) or by analyzing the variance of client updates.

A concept from cooperative game theory used to quantify the marginal contribution of each client's data to the global model's performance.

• Role in Bias Mitigation: Provides a principled, data-driven method for client valuation. This can identify clients whose data is uniquely valuable for improving fairness or accuracy on underrepresented classes.
• Application: Clients with high Shapley values for improving performance on a minority class can be prioritized in selection to mitigate performance bias.
• Computational Note: Exact calculation is combinatorially expensive; efficient approximations (e.g., TMC-Shapley) are used in practice.

A small, weighted subset of client data (or synthetic proxy data) that approximates the overall federated data distribution.

• Role in Bias Mitigation: Serves as a bias diagnostic and corrective tool.
• Diagnostic: The server can use a coreset to estimate global data distribution and identify underrepresented patterns.
• Corrective: The selection policy can be designed to choose clients whose local data best complements the current coreset, actively filling representation gaps.
• Advanced Use: In some protocols, the coreset itself is used to compute proxy gradients on the server, guiding bias-aware client selection.