Decentralized Orchestration Architecture for Fluid Computing: A Secure Distributed AI Use Case

A team of researchers from the University of Vigo and other institutions has published a paper proposing a novel Decentralized Orchestration Architecture designed for the emerging paradigm of Fluid Computing. The core challenge they address is managing distributed AI and IoT applications across heterogeneous resources—spanning end devices, edge nodes, and cloud platforms—that are often under separate administrative control. Current solutions tend to be centralized and struggle with multi-domain coordination. Their architecture introduces an "orchestration plane" that allows different domains to coordinate deployments in a decentralized manner while preserving local autonomy, enabling intent-based service placement across the entire computing continuum.

The paper validates the architecture with a critical use case: securing Decentralized Federated Learning (DFL) against Byzantine threats, where malicious participants can submit corrupted data or models. The researchers leverage domain-side control capabilities to implement FU-HST, a Software-Defined Networking (SDN)-enabled, multi-domain anomaly detection mechanism. This system works alongside traditional Byzantine-robust aggregation methods to identify and mitigate attacks. They tested the approach via simulation in both single- and multi-domain settings, evaluating its effectiveness in anomaly detection, its impact on DFL model performance, and the associated computation and communication overhead. The work, currently under peer review and available on arXiv, represents a significant step toward making large-scale, cross-organizational AI deployments both feasible and secure.