[Paper Review] FederatedTrust: A Solution for Trustworthy Federated Learning

Pedro Miguel Sánchez Sánchez, Alberto Huertas Celdrán+4 more|arXiv (Cornell University)|Feb 20, 2023

Privacy-Preserving Technologies in Data4 citations

TL;DR

FederatedTrust proposes a comprehensive trustworthiness taxonomy for Federated Learning (FL) comprising six pillars—privacy, robustness, fairness, explainability, accountability, and federation—along with over 30 metrics. It introduces an algorithm and prototype integrated into FederatedScope that computes a holistic trust score, validated across five experiments using FEMNIST and N-BaIoT datasets, demonstrating its utility in assessing FL model trustworthiness under diverse configurations.

ABSTRACT

The rapid expansion of the Internet of Things (IoT) and Edge Computing has presented challenges for centralized Machine and Deep Learning (ML/DL) methods due to the presence of distributed data silos that hold sensitive information. To address concerns regarding data privacy, collaborative and privacy-preserving ML/DL techniques like Federated Learning (FL) have emerged. However, ensuring data privacy and performance alone is insufficient since there is a growing need to establish trust in model predictions. Existing literature has proposed various approaches on trustworthy ML/DL (excluding data privacy), identifying robustness, fairness, explainability, and accountability as important pillars. Nevertheless, further research is required to identify trustworthiness pillars and evaluation metrics specifically relevant to FL models, as well as to develop solutions that can compute the trustworthiness level of FL models. This work examines the existing requirements for evaluating trustworthiness in FL and introduces a comprehensive taxonomy consisting of six pillars (privacy, robustness, fairness, explainability, accountability, and federation), along with over 30 metrics for computing the trustworthiness of FL models. Subsequently, an algorithm named FederatedTrust is designed based on the pillars and metrics identified in the taxonomy to compute the trustworthiness score of FL models. A prototype of FederatedTrust is implemented and integrated into the learning process of FederatedScope, a well-established FL framework. Finally, five experiments are conducted using different configurations of FederatedScope to demonstrate the utility of FederatedTrust in computing the trustworthiness of FL models. Three experiments employ the FEMNIST dataset, and two utilize the N-BaIoT dataset considering a real-world IoT security use case.

Motivation & Objective

• To address the growing need for trustworthiness in FL beyond data privacy, especially in real-world deployments.
• To identify and formalize trustworthiness pillars and metrics specific to FL, extending beyond existing AI trust frameworks.
• To design and implement a flexible, extensible algorithm—FederatedTrust—that computes a holistic trust score for FL models.
• To validate the prototype within the FederatedScope FL framework across diverse configurations, including differential privacy and client selection strategies.
• To demonstrate the feasibility and utility of systematic trust assessment in FL using empirical experiments with real-world datasets.

Proposed method

• Developed a six-pillar taxonomy for trustworthy FL: privacy, robustness, fairness, explainability, accountability, and federation, each with specific sub-notions and metrics.
• Defined over 30 quantifiable metrics across the pillars, including client reputation, selection fairness, model accuracy, and data imbalance indicators.
• Designed the FederatedTrust algorithm to compute a global trust score by aggregating individual metric scores using configurable weights.
• Implemented the algorithm as a prototype integrated into the FederatedScope FL framework to support real-time trust assessment during training.
• Employed a weighted aggregation strategy for metrics, allowing dynamic configuration based on deployment context and stakeholder priorities.
• Conducted five experiments with varying client counts, selection rates, training rounds, and differential privacy parameters to evaluate trust score dynamics.

Experimental results

Research questions

• RQ1 How can trustworthiness in Federated Learning be systematically evaluated beyond data privacy?
• RQ2 What are the most relevant trustworthiness pillars and metrics specifically applicable to FL models?
• RQ3 Can a unified, configurable framework effectively compute a holistic trust score for FL models across diverse configurations?
• RQ4 How do different FL configurations (e.g., client selection, differential privacy) affect the computed trustworthiness levels?
• RQ5 To what extent can the FederatedTrust algorithm detect and reflect trust-related issues such as bias, data imbalance, or model drift?

Key findings

• The FederatedTrust prototype successfully computed trust scores across diverse FL configurations, demonstrating its adaptability and configurability.
• Experiments revealed that differential privacy and client selection fairness significantly impacted the overall trust score, highlighting trade-offs in model design.
• The federation pillar—assessing both participant and model trust—provided unique insights into system-level reliability and cooperation dynamics.
• The trust score was sensitive to data imbalance and model drift, with metrics like client participation variation and model accuracy showing strong correlation with trust levels.
• The algorithm detected performance degradation in models trained with non-iid data, especially when combined with differential privacy, indicating its sensitivity to real-world FL challenges.
• The prototype’s integration into FederatedScope enabled real-time monitoring of trustworthiness, proving its feasibility in production-like environments.