Federated Learning (FL) allows multiple clients—such as different hospitals or regional banks—to collaboratively train a machine learning model without sharing their raw, private data. As these models are deployed in high-stakes environments, ensuring they do not discriminate against protected groups (based on race, gender, etc.) is critical.
However, measuring fairness in an FL setting introduces a unique challenge. Fairness can be evaluated in two distinct ways:
- Global Fairness: Is the model fair when evaluated across the entire combined dataset of all clients?
- Local Fairness: Is the model fair when evaluated on an individual client's specific local dataset?
Because data is often distributed unevenly across clients (data heterogeneity), local and global fairness can clash. For example, one bank might serve a predominantly minority neighborhood while another does not. Prior work usually focuses on achieving either local or global fairness without a clear understanding of how they interact. This paper asks a fundamental question: When does local fairness imply global fairness? And what are the absolute mathematical limits of trying to achieve both alongside high accuracy?
An Information-Theoretic Approach to Fairness
The authors frame fairness through the lens of Information Theory, specifically using Mutual Information to measure how much a model's prediction ($\hat{Y}$) depends on a sensitive attribute ($Z$).
- Global Disparity is defined as $I(Z; \hat{Y})$.
- Local Disparity is defined as $I(Z; \hat{Y} | S)$, which is the mutual information conditioned on the client $S$.
The core novelty of this work is applying Partial Information Decomposition (PID) to split the total disparity into three non-negative components. This decomposition mathematically proves why achieving fairness at the local client level doesn't guarantee global fairness, and vice versa.
As illustrated above, PID breaks down the relationship between global and local disparity into three distinct regions:
- Unique Disparity (Agreement Region): Unfairness coming solely from the model's predictions, independent of the client distribution. Here, local and global fairness align.
- Redundant Disparity (Disagreement Region): Unfairness that overlaps between the model's predictions and the client's identity (e.g., a specific client mostly serves one demographic).
- Masked or Synergistic Disparity (Disagreement Region): Unfairness that only appears when you look at the prediction and the client identity together.
The Danger of Naive Mitigation
By breaking down disparity into these three components, the authors reveal critical insights into the behavior of FL models:
- Local Fairness $\neq$ Global Fairness: If Redundant Disparity exists (e.g., demographics are heavily skewed across clients), applying local fairness constraints at each client will not result in a globally fair model.
- Global Fairness $\neq$ Local Fairness: Conversely, a globally fair model can still be locally unfair if Masked Disparity exists. For example, if a model unfairly favors men at Client A and women at Client B, the biases might cancel out globally, but the model remains highly discriminatory at the local level.
Because of these hidden disparities, applying standard fairness regularizers locally without understanding the underlying data distribution can actually exacerbate global disparities.
Hard Mathematical Ceilings: The AGLFOP Framework
To understand the absolute limits of mitigating these disparities, the authors introduce the Accuracy and Global-Local Fairness Optimality Problem (AGLFOP). This convex optimization framework evaluates all potential joint distributions to find the absolute Pareto frontier—the theoretical ceiling of accuracy given strict global and local fairness constraints.
The researchers validated their framework on both custom synthetic datasets (designed to explicitly control statistical dependence) and the real-world Adult (UCI) dataset. Key empirical results include:
- IID Data: In scenarios with uniformly distributed data, Unique Disparity dominated. Local and global fairness aligned perfectly.
- Heterogeneous Data: When data was split with high demographic skew (e.g., Client 0 is 90% women, Client 1 is 90% men), Redundant Disparity dominated. The resulting model was locally fair but globally unfair.
- Synergistic Data: When true labels aligned with an XOR of the sensitive attribute and client ID, Masked Disparity dominated, creating a globally fair but locally unfair model.
- Pareto Frontiers: Using AGLFOP on the Adult dataset, the authors demonstrated that you cannot simply optimize your way out of a bad data distribution. Under high demographic heterogeneity, demanding perfect local and global fairness caps the maximum possible model accuracy at roughly 64%.
Limitations and Future Directions
While PID offers a powerful theoretical lens, estimating PID terms is computationally expensive. As the number of clients or sensitive attributes grows, the computational cost of the convex optimization scales up significantly, and the accuracy relies heavily on the quality of empirical estimators.
Future work will explore more efficient PID computation techniques, such as neural estimation. Additionally, while AGLFOP currently focuses on statistical parity (independence between $Z$ and $\hat{Y}$), the framework could be adapted to explore other fairness metrics like equalized odds, or modified to penalize only the worst-case client fairness scenarios.
Summary updated