GlobLearn: Exploring Privacy – Preserving Federated Learning for Supply Chain Collaboration

Globis publishes research white paper on federated learning approach enabling collaborative AI without sharing confidential data.

The logistics industry sits on a goldmine of data that could power transformative AI predictions — but competitive concerns and privacy regulations keep it locked in silos. What if supply chain partners could collectively train predictive models without any confidential data ever leaving their premises? 

That’s the question the ZEE research project (Zero-Data Exchange for Engineering) set out to answer. Over the past three years, our consortium — Globis, Ahlers Belgium, Nallian, and Sirris — investigated how federated learning could break through the data sharing paradox in logistics. 

The Challenge 

A single shipment generates data across 5–10 organisations, yet no party sees the complete picture. Current ETA prediction systems typically achieve only 60–70% accuracy, relying on static calculations rather than learning from patterns across diverse datasets. Machine learning could dramatically improve these predictions — but centralising data from competing logistics companies is simply not feasible. 

The GlobLearn Research Platform 

GlobLearn is the reference architecture and research platform we developed to explore this challenge. Built on the principle of “Zero Exchange,” the platform demonstrates that supply chain partners can collectively train predictive models while maintaining complete data sovereignty. Only model parameters — never raw data — cross organisational boundaries. 

A key research contribution is the three-plane architecture we developed: a Management Plane for use case definition and participant management, a Control Plane for federated learning orchestration with integrated data quality validation, and a Data Plane with ephemeral processing environments at each participant’s premises.

What We Learned 

The research validated several important findings:  

• Federated learning can be applied to logistics use cases despite highly heterogeneous data across carriers 

• Integrated data quality gates are essential — without them, participants with poor data degrade the model for everyone 

• Ephemeral data environments provide practical privacy guarantees beyond the federated learning protocol itself 

• For ETA forecasting, proof-of-concept experiments achieved RMSE values of 5.7–5.9 hours, demonstrating meaningful prediction capability from distributed data 

We also encountered significant challenges: building trust with potential participants, managing data quality variability across diverse legacy systems, and the engineering complexity of integrating multiple open-source frameworks into a coherent system.

What’s Next 

The ZEE project has laid an architectural and conceptual foundation. Significant further research and development would be needed to bring this concept to production readiness — including validation with truly independent participants, scalability testing, and security certification. We’re exploring options for a follow-up trajectory to address these challenges. 

📄 Download the Research White Paper (PDF)

About the ZEE Project 

The ZEE project (HBC.2022.0423) was a 36-month collaborative research initiative running from January 2023 to December 2025, funded under the VLAIO Onderzoeksproject programme as part of the ITEA framework. The consortium consisted of Globis (platform development), Ahlers Belgium (breakbulk logistics), Nallian (air cargo logistics), and Sirris (research partner). 

This research was supported by VLAIO (Flanders Innovation & Entrepreneurship) and funded by the European Union – NextGenerationEU.