ENFIELD research on Trustworthy Federated Learning presented at MIST 2026
FedXParity framework brings fairness, privacy and explainability together in federated learning.
A new research addressing some of the key challenges surrounding trustworthy artificial intelligence was presented at the 31st European Symposium on Research in Computer Security (ESORICS 2026), held from 14 to 18 September at Sapienza University of Rome.
The paper, “FedXParity: A Unified Framework for Fairness, Privacy, and Explainability in Federated Learning,” was presented by Aida Akbarzadeh from NTNU on Mobile and IoT Security Technology (MIST), organised in conjunction with ESORICS 2026.
Authored by Reza Shahbazian, Georgios Spathoulas and Aida Akbarzadeh, the research is one of the outcomes of a TES research exchange involving Reza Shahbazian and contributes to ENFIELD’s broader work towards developing trustworthy and responsible artificial intelligence.
Trustworthy AI and Federaterd Learning
Federated Learning enables multiple organisations, systems or devices to collaboratively train artificial intelligence models without having to centralise their underlying data. This makes it particularly relevant in environments where privacy, security or data ownership are critical. However, protecting the data is only one part of the challenge.
Federated Learning systems must also address questions of fairness, transparency and explainability, particularly when models operate across heterogeneous devices, users or environments.
FedXParity proposes a unified approach designed to address four dimensions simultaneously: predictive performance, algorithmic fairness, privacy and explainability.
A distinctive aspect of the framework is its focus on explanatory fairness. Rather than considering only whether a model produces equitable outcomes, the research also examines whether the reasoning and features behind those decisions are treated consistently across different groups.
This issue can be particularly important in areas such as Cyber-Physical Systems, Internet of Things-enabled environments and enterprise mobile infrastructures, where biased model reasoning could make vulnerabilities or systemic differences harder to identify.
Promising results for fairer and more privacy-aware AI
FedXParity introduces a local regularisation mechanism that aligns feature-attribution distributions across sensitive groups. In practical terms, the approach aims to reduce differences not only in model outcomes, but also in how models arrive at those outcomes.
The framework was evaluated using the ACS Employment dataset as well as a synthetic Cyber-Physical Systems power-grid substation dataset, including experiments under challenging Non-IID data conditions commonly encountered in federated environments.
The results show that FedXParity was able to reduce the Equalized Odds gap by up to 75%, while also mitigating privacy leakage.
Importantly, the research indicates that these privacy benefits can be achieved without relying exclusively on traditional differential privacy techniques, which can introduce significant trade-offs in model utility.
By considering fairness, privacy and explainability together, rather than treating them as separate optimisation problems, FedXParity contributes to ongoing research into AI systems that are not only effective, but also more transparent, equitable and privacy-aware.
ENFIELD on a leading European security forum
Established in 1990, the European Symposium on Research in Computer Security (ESORICS) provides an international forum for research in computer security, bringing together researchers, system developers and related research communities.
Its 31st edition took place at the Main Campus of Sapienza University of Rome from 14 to 18 September 2026, combining the main scientific programme with a series of specialised workshops covering emerging challenges in cybersecurity, privacy and artificial intelligence.
The presentation of FedXParity at MIST highlights the role of research exchanges and cross-institutional collaboration within ENFIELD in generating new research at the intersection of artificial intelligence, cybersecurity and trustworthy AI.
Through these collaborations, ENFIELD continues to support the development and transfer of knowledge needed to build AI technologies that can be deployed responsibly in increasingly complex and security-sensitive environments.
Paper information
Title: FedXParity: A Unified Framework for Fairness, Privacy, and Explainability in Federated Learning
Authors: Reza Shahbazian, Georgios Spathoulas and Aida Akbarzadeh
Read more: FedXParity on ENFIELD’s ZENODO Community
