🤖 machine learning

Beyond performance-wise Contribution Evaluation in Federated Learning

This paper argues that current federated learning evaluation methods are insufficient because they focus solely on performance metrics, proposing instead a comprehensive Shapley value-based framework to quantify client contributions across independent dimensions of trustworthiness—reliability, resilience, and fairness—to enable more equitable reward allocation.

Original authors: Balazs Pejo

Published 2026-02-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Balazs Pejo

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 the captain of a massive ship, and you've hired a crew of 20 different specialists to help you navigate. Your goal is to build the best possible map for the journey. In the world of Artificial Intelligence, this is called Federated Learning. Instead of everyone bringing their private maps to a central office (which would be a privacy nightmare), they each work on their own piece of the map and send you updates.

For years, the only way to decide who gets a bonus or who gets fired was to ask one simple question: "How accurate is your piece of the map?"

If a crew member's map section was 99% correct, they got a gold star. If it was 90%, they got a smaller star. The paper you provided argues that this "Gold Star" system is dangerously flawed. It's like judging a chef only by how fast they can chop vegetables, ignoring whether the food tastes good, if it's safe to eat, or if they are being fair to all the ingredients.

Here is the breakdown of the paper's findings using a simple, everyday analogy:

The Four Pillars of a "Good" Crew Member

The authors say we need to evaluate our crew members on four different things, not just one. They call these the pillars of Trustworthiness:

Performance (The Speed): How accurate is the map? (This is what everyone currently measures).
Fairness (The Equity): Does the map treat everyone equally? Imagine if a map was great for rich neighborhoods but completely wrong for poor ones. That's unfair. A "Fair" crew member ensures the map works for everyone, regardless of who they are.
Reliability (The Sturdiness): What happens if it starts raining or the paper gets smudged? A "Reliable" crew member provides a map that still makes sense even if the data is a little noisy or messy.
Resilience (The Defense): What if a saboteur tries to trick the map? A "Resilient" crew member provides a map that can't be easily fooled by someone trying to trick the system (like a hacker trying to make a stop sign look like a speed limit sign).

The Big Surprise: The "Jack of All Trades" Myth

The most shocking discovery in this paper is that no single crew member is good at all four things.

In fact, the paper found that these skills are completely independent.

You might have a crew member who is a genius at accuracy (Performance) but their map is terrible at being fair (Fairness). They might accidentally create a map that only works for one specific group of people.
You might have another crew member who is mediocre at accuracy but their map is incredibly sturdy (Reliability). Even if the data is messy, their map doesn't fall apart.
You might have a third who is great at defense (Resilience) but harmful to fairness.

The Analogy:
Think of it like a sports team.

Player A is the best scorer in the league (High Performance), but they are a terrible teammate who never passes the ball and gets the whole team disqualified for unsportsmanlike conduct (Low Fairness).
Player B is a defensive wall who stops every attack (High Resilience), but they never score a goal (Low Performance).

If you only look at the scoreboard (Performance), you will reward Player A and fire Player B. But if your goal is to win the championship (which requires teamwork, fairness, and defense), you are making a huge mistake.

The "Shapley Value" (The Fairness Calculator)

To figure out who deserves what, the authors used a mathematical tool called the Shapley Value. Think of this as a very sophisticated calculator that asks: "If we remove this specific person from the team, how much does the team's overall quality drop?"

They ran this calculator four times for every person:

How much did they help the score?
How much did they help fairness?
How much did they help reliability?
How much did they help resilience?

The Results: A Wake-Up Call

When they compared the results, they found that the rankings were completely different depending on which question you asked.

The person who was #1 for Accuracy might be #20 for Fairness.
The person who was #1 for Resilience might be #15 for Accuracy.

This means that if you only pay people based on accuracy, you are accidentally punishing the people who are keeping the system safe, fair, and robust. You might be rewarding the person who is building a "fast but broken" system.

Why Does This Matter?

The authors argue that in the real world, we can't just care about speed or accuracy.

In Healthcare: A model that is 99% accurate but discriminates against a specific race is useless and dangerous.
In Self-Driving Cars: A car that drives fast but crashes when it sees a weird shadow (low reliability) is a death trap.
In Finance: A system that makes money but gets hacked easily (low resilience) will lose everything.

The Takeaway

The paper concludes that we need to stop using a "one-size-fits-all" ruler to measure our AI teams. We need a multi-dimensional report card.

If we want to build AI that we can actually trust, we need to reward the "Defenders" and the "Fairness Champions" just as much as the "Speedsters." Otherwise, we are building systems that are fast, but fragile, unfair, and easily broken.

1. Problem Statement

Federated Learning (FL) enables collaborative model training without sharing raw data, but its success relies on fair and accurate evaluation of participant contributions.

Current Limitation: Existing client evaluation methods predominantly rely on a single performance metric, typically accuracy or loss.
The Gap: This performance-centric view is insufficient for building Trustworthy AI. A client might improve model accuracy while simultaneously degrading the model's fairness, reliability (tolerance to noise), or resilience (resistance to adversarial attacks).
Core Question: How can we quantify client contributions across multiple, independent dimensions of trustworthiness, and do these dimensions correlate with traditional performance metrics?

2. Methodology

The authors propose a multifaceted evaluation framework that extends client contribution analysis beyond accuracy to include three pillars of trustworthiness: Fairness, Reliability, and Resilience.

A. Evaluation Metrics

The paper defines four distinct scoring functions ( $v$ ) to evaluate the global model $M$ :

Performance ($perf$): Standard classification accuracy.
Fairness ($fair$): Measured via Demographic Parity. It calculates the difference in prediction probabilities between a sensitive group and the rest of the population. The score is $1 - |P(\hat{y}|S) - P(\hat{y}|S^c)|$ , where higher values indicate better fairness.
Reliability ($rel$): Measured by the model's stability against random noise (Gaussian noise). It calculates the probability that the model's prediction remains unchanged after perturbation.
Resilience ($res$): Measured by the model's resistance to adversarial attacks (specifically Projected Gradient Descent - PGD). It calculates the success rate of the attack on correctly classified samples, inverted so higher values mean better resilience.

B. Contribution Estimation Algorithms

To attribute value to individual clients based on these metrics, the authors employ two Shapley Value approximation methods:

Leave-One-Out (LOO): A lightweight heuristic where a client's contribution is the difference in the utility score when the client is present versus absent from the coalition.
- $LOO(k) = v([K]) - v([K] \setminus \{k\})$
Guided-Truncated-Gradient (GTG): A state-of-the-art approximation of the Shapley Value. It optimizes computation by:
- Skipping rounds with negligible global improvement.
- Sampling a subset of client permutations.
- Ignoring marginal contributions below a specific threshold.

C. Experimental Setup

Datasets: Three distinct domains were used to ensure generalizability:
- Tabular: Adult (MLP model).
- Computer Vision: CIFAR-10 (CNN model).
- NLP: IMDB (USE encoder).
Settings: Experiments were conducted under both IID (Independent and Identically Distributed) and Non-IID (using Dirichlet distribution $\alpha=0.5$ ) data distributions.
Client Counts: Simulations with $K=4$ and $K=20$ clients.
Implementation: Built using the Flower framework and TensorFlow.

3. Key Contributions

Multidimensional Evaluation: The paper is the first to systematically evaluate FL client contributions across four distinct axes: Performance, Fairness, Reliability, and Resilience.
Demonstration of Independence: The authors provide empirical evidence that these trustworthiness dimensions are statistically uncorrelated. A client ranked highly for accuracy may be ranked poorly (or even negatively) for fairness or resilience.
Methodological Validation: The findings hold true across different valuation methodologies (LOO vs. GTG), proving the robustness of the conclusion.
Cross-Domain Generalizability: The study validates these findings across tabular, image, and text data modalities.

4. Key Results

Divergent Rankings: Client rankings shift dramatically depending on the metric used. Clients who are "high performers" (high accuracy) can be "detrimental" to fairness or resilience.
Low Correlation with Performance:
- Spearman correlation coefficients ( $\Phi$ ) between performance scores and trustworthiness scores are consistently low (often < 0.1) or even negative.
- This indicates that optimizing for accuracy does not guarantee improvements in other trustworthiness dimensions.
Inter-correlation of Trustworthiness:
- While independent from performance, Reliability and Resilience show strong positive correlation (often > 0.5), as both measure robustness.
- In the Adult dataset, Fairness showed a high correlation with robustness metrics, suggesting a link between equitable data representation and resistance to perturbations.
Distribution Shifts: Under Non-IID settings, trustworthiness scores (especially for LOO) showed distinct negative or positive shifts compared to the symmetric, zero-centered distributions seen in IID settings.

5. Significance and Implications

Flaw in Current Incentive Schemes: The paper argues that current FL reward systems based solely on accuracy are fundamentally flawed. They may inadvertently penalize clients who provide critical data for fairness or robustness, leading to suboptimal and potentially unethical global models.
Need for Holistic Governance: To build Trustworthy AI (aligned with the EU AI Act), consortiums must adopt multi-dimensional evaluation schemes. A single metric is inadequate for equitable reward allocation.
Future Directions: The authors suggest that future work should focus on integrating these multi-dimensional scores into a unified reward mechanism and exploring how explicit training techniques (e.g., fairness regularization, adversarial training) affect contribution scores.

Conclusion

The paper fundamentally challenges the "performance-only" paradigm in Federated Learning. It demonstrates that trustworthiness is multi-faceted and client-dependent. A comprehensive evaluation framework is essential to ensure that collaborative AI systems are not only accurate but also fair, reliable, and resilient.