Cyberwal in Galaxia
Academy

Earth Observation (EO) data are increasingly critical in sensitive domains such as defense, border surveillance, and critical infrastructure monitoring. These datasets are not only voluminous (e.g., multispectral to hyperspectral and SAR imagery) but also frequently classified, sovereign, and geographically dispersed across satellites, ground stations, cloud, edge, and HPC infrastructures. This fragmentation introduces significant cybersecurity risks: data interception, unauthorized access through API vulnerabilities, AI model attacks (e.g., backdoors, gradient inversion), and leakage via insecure multi-tier storage or misconfigured services. To address these challenges, this talk presents the design and practical implementation of Federated Learning (FL) within EarthLab’s Max-ICS platform (https://www-max-ics.earthlab.lu/), enabling secure, distributed EO data processing in sensitive environments. FL allows multiple stakeholders, such as defense ministries, space agencies, and interoperable coalitions (e.g., NATO, EU), to collaboratively train AI models without centralizing sensitive data, preserving sovereignty and reinforcing trust across organizations. Beyond data decentralization, we also explore how Homomorphic Encryption (HE) can be integrated into the FL workflow to ensure end-to-end confidentiality of exchanged model parameters and gradients. By allowing computations directly on encrypted data, HE strengthens privacy guarantees, preventing information leakage during model aggregation or communication phases. The presentation will discuss the architectural integration of HE into Max-ICS, the trade-offs between performance and security, and lessons learned from practical testing. We will conclude with a roadmap toward resilient, scalable, and trustworthy AI architectures for secure Earth Observation ecosystems.