🤖 machine learning

FedCF: Fair Federated Conformal Prediction

This paper introduces FedCF, a framework that extends Conformal Fairness to the Federated Learning setting to audit and ensure equitable uncertainty quantification across diverse demographic groups while leveraging the exchangeability assumption.

Original authors: Anutam Srinivasan, Aditya T. Vadlamani, Amin Meghrazi, Srinivasan Parthasarathy

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anutam Srinivasan, Aditya T. Vadlamani, Amin Meghrazi, Srinivasan Parthasarathy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a team of chefs from different neighborhoods to create a single, perfect recipe book. Each chef has their own local ingredients and cooking style, but they can't share their secret recipes or ingredient lists with each other due to privacy rules. This is the world of Federated Learning: many people training a model together without sharing their private data.

The problem? Sometimes, the final recipe book might taste great for some groups of people but be terrible for others. Maybe it works perfectly for spicy food lovers but fails miserably for those who prefer mild flavors. In machine learning terms, this is unfairness.

This paper introduces a new tool called FedCF (Federated Conformal Fairness) to fix this. Here is how it works, broken down into simple concepts:

1. The "Safety Net" (Conformal Prediction)

Before we talk about fairness, we need to understand the "Safety Net." In machine learning, models usually just guess an answer (e.g., "This is a cat"). But what if the model is unsure?

Standard Prediction: "This is a cat." (If wrong, you have no warning).
Conformal Prediction (CP): "I am 95% sure this is a cat, but it could also be a dog." It gives you a list of possibilities. If the model is confident, the list is short (just "cat"). If it's unsure, the list is long ("cat, dog, fox").
The Guarantee: The paper guarantees that if you say "95% confidence," the real answer will actually be in that list 95% of the time. It's a mathematical promise of reliability.

2. The Problem: The Safety Net is Biased

The authors found that while this "Safety Net" works well on average, it's not fair.

The Analogy: Imagine the Safety Net is a fishing net. For the "rich" group of data, the net has small holes, catching almost everything. For the "poor" group, the net has huge holes, letting many fish escape.
The Result: The model might be 95% reliable for Group A, but only 80% reliable for Group B. This is unfair, especially in critical fields like healthcare or finance.

3. The Solution: FedCF (The Fairness Auditor)

FedCF is a new way to check and fix this bias without the chefs (clients) ever showing their secret ingredients (data) to the head chef (server).

Here is the step-by-step process:

The "Local Audit": Each client (chef) looks at their own local data. They count how many times their "Safety Net" catches the right answer for different groups (e.g., "How often does my net catch the right answer for Group A vs. Group B?"). They don't send the data; they just send a simple number (a score) representing this count.
The "Global Puzzle": The server collects these simple numbers from everyone. It doesn't know who has what data, but it can mathematically combine these numbers to figure out the overall fairness of the whole system.
The "Adjustment Knob": If the server sees that Group B's net is too loose (unfair), it turns a "knob" (a threshold value). This tightens the net for Group B, making the prediction list slightly larger (less efficient) but more accurate.
The Descent Search: Instead of guessing the knob setting one by one (which takes forever), FedCF uses a smart "descent" method. It's like sliding down a hill to find the lowest point. It quickly finds the perfect setting where the net is fair for everyone, without needing thousands of rounds of communication.

4. Two Ways to Play (Privacy vs. Speed)

The paper offers two ways to send these numbers, depending on what you care about more:

The "Speedy" Way: Clients send just a few numbers. It's fast and uses less internet bandwidth, but a sneaky server might be able to guess a little bit about the data distribution.
The "Private" Way: Clients do more math on their own and send a "difference" number (like saying "I have 5 more than average" instead of "I have 5"). This hides the individual data much better, but it requires sending slightly more information.
The Hybrid: You can mix and match! Some clients can be speedy, others can be private, and the system still works.

5. What They Found (The Results)

The authors tested this on real-world data (like income records, education levels, and skin condition images) split across different regions.

The Good News: FedCF successfully made the "Safety Net" fair. It ensured that the prediction reliability was roughly the same for all groups (e.g., if it was 95% for one group, it was 95% for the other).
The Trade-off: To make it fair, the "lists" of possibilities sometimes got a tiny bit longer (less efficient). But the paper shows this cost is small compared to the benefit of fairness.
The "Audit" Feature: They also showed that this system can be used as a "report card." Regulators can use it to check if a model is fair without needing to see the private data of the people involved.

Summary

FedCF is like a fairness inspector for a team of remote workers. It ensures that the final product (the AI model) treats everyone equally, even though the workers never shared their private notes. It uses smart math to adjust the model's confidence levels so that no group gets a "leaky" safety net, all while respecting privacy and keeping communication fast.

Technical Summary: FedCF: Fair Federated Conformal Prediction

Problem Statement

Machine learning models often exhibit disparate performance across different demographic subgroups, leading to biases that can be exacerbated in safety-critical domains like healthcare and finance. While Conformal Prediction (CP) provides rigorous, model-agnostic uncertainty quantification with finite-sample coverage guarantees, standard CP is agnostic to sensitive attributes and does not ensure fairness. Existing "Fair Conformal Prediction" (CF) methods address this by ensuring conditional coverage guarantees across subgroups, but they rely on centralized data access.

In Federated Learning (FL), data remains decentralized to preserve privacy, making the application of standard CF impossible. Furthermore, recent work indicates that shifting toward privacy-preserving decentralized training can inadvertently exacerbate unfairness due to data heterogeneity and varying client capabilities. The core challenge is to extend the theoretical guarantees of CF to a federated setting without violating privacy constraints or incurring prohibitive communication overhead, while accounting for the non-IID nature of federated data.

Methodology

The authors propose FedCF (Federated Conformal Fairness), a framework that adapts the Conformal Fairness (CF) framework to the federated setting. The methodology rests on three pillars:

1. Descent-Based Optimization Reformulation

The original CF algorithm employs a grid search over a discretized space to find the minimal threshold $\lambda$ satisfying fairness constraints. This is inefficient in FL due to the high cost of communication rounds. FedCF reformulates this as a descent-based optimization problem. Instead of a grid search, the algorithm adaptively updates the threshold $\lambda$ using a momentum-based update rule. This allows the system to converge to a constraint-satisfying $\lambda$ in significantly fewer communication rounds (demonstrated as a ~10x speedup in experiments) while escaping local minima via random restarts.

2. Mixture-of-Clients Coverage Gap Formulation

To handle decentralized data, FedCF cannot directly estimate the conditional coverage required for fairness metrics. The authors introduce a mixture-of-clients formalization based on the assumption of partial exchangeability. They decompose the global fairness-specific coverage level ( $\Pi_{cov}$ ) into four terms:

Terms I & II: Computed locally by each client based on their calibration data and the specific group-label pair.
Term III: The probability that a test point is exchangeable with a specific client's data (weighted by client size).
Term IV: The global probability of the fairness filter being active.

By deriving interval bounds (Theorem 3.1) for these terms, the server can aggregate local statistics to bound the global coverage gap without ever seeing raw client data. The framework also offers a Maximum Likelihood Estimation (MLE) variant for tighter, though less conservative, estimates under a partial-IID assumption.

3. Flexible Aggregation Protocols

FedCF addresses the trade-off between communication efficiency and data privacy through two protocols:

Communication Efficient: Clients send summary statistics ( $\alpha$ values) for each group-label pair. The server aggregates these to compute bounds. This has linear communication complexity relative to the number of groups.
Enhanced Privacy: Clients compute pairwise differences of coverage terms locally before sending them to the server. This obfuscates individual distributional information, preventing reconstruction attacks, at the cost of quadratic communication complexity relative to the number of groups.
Hybrid: Clients can choose between protocols based on their resource constraints.

The framework also supports Differential Privacy (DP) by adding Gaussian noise to client updates, ensuring robustness against malicious servers.

Key Contributions

Extension of CF to FL: The paper presents the first framework to extend Conformal Fairness to a federated setting, theoretically proving that conditional coverage guarantees can be maintained in a distributed environment under partial exchangeability.
Descent-Based Algorithm: A novel optimization approach that replaces grid search with an adaptive descent method, drastically reducing the number of communication rounds required for convergence in FL.
Theoretical Bounds: The derivation of interval bounds for fairness-specific coverage levels in a federated setting, providing a rigorous method to audit fairness gaps without centralizing data.
Privacy-Communication Trade-offs: The proposal of distinct aggregation protocols (Communication Efficient, Enhanced Privacy, and Hybrid) allowing clients to balance bandwidth constraints with privacy requirements.
Empirical Validation: Extensive experiments on tabular (ACSIncome, ACSEducation), graph (Pokec), and vision (Fitzpatrick) datasets demonstrate that FedCF successfully controls fairness disparities (e.g., Demographic Parity, Equal Opportunity) while maintaining the original CP coverage guarantees.

Experimental Results

The authors evaluated FedCF across multiple domains and fairness metrics:

Fairness Control: FedCF consistently achieves the user-specified closeness criterion ( $c$ ) for fairness disparity, whereas baseline Federated Conformal Prediction (FCP) without fairness constraints fails to control these gaps.
Efficiency vs. Fairness: There is an inherent trade-off; enforcing stricter fairness (smaller $c$ ) increases the prediction set size (reducing efficiency). However, FedCF minimizes this cost compared to baselines.
Robustness to Heterogeneity: The framework performs effectively across varying numbers of clients (2 to 48) and different data partitioning schemes (natural geographic splits vs. Dirichlet partitions).
Graph Data: FedCF successfully integrates with graph-specific non-conformity scores (DAPS), demonstrating applicability beyond standard tabular data.
Privacy Protocols: Experiments show that while the Enhanced Privacy protocol is more conservative (larger prediction sets) due to tighter bounds, it effectively controls fairness disparities while protecting client data.

Significance and Claims

The paper claims that FedCF provides a theoretically grounded and practically viable solution for auditing and ensuring fairness in federated machine learning models. Its significance lies in:

Regulatory Compliance: Offering a mechanism to satisfy emerging regulations (e.g., EU AI Act, NYC Local Law 144) that require fairness audits in decentralized systems without compromising data privacy.
Trustworthy ML: Bridging the gap between uncertainty quantification and fairness, ensuring that models not only predict with known error bounds but do so fairly across sensitive groups.
Auditability: The framework can be used post-hoc to audit existing federated models, determining if they meet fairness specifications without requiring retraining or access to raw data.

The authors acknowledge limitations, specifically the reliance on the partial exchangeability assumption, which may be violated in settings with dynamic data drift. They suggest future work involving non-exchangeable CP and robust federated learning techniques to address these scenarios.