Distributed Cloud-Edge-IoT systems require accurate knowledge of node
state to enable adaptive orchestration and resource management. However, exchanging full-state information in decentralized environments incurs high communication overhead, imposing increasing limitations as the system scales.
This paper proposes a communication-efficient framework for decentralized full-state estimation based on autoencoder-derived embeddings. The approach enables nodes to exchange compact representations instead of raw state vectors while preserving reconstruction fidelity.
We introduce a tolerance-based evaluation metric aligned with operational requirements and perform a systematic analysis under varying embedding dimensions, node populations, domain-shift scenarios, and gossip-based dissemination. Results show that
autoencoder embeddings consistently outperform linear PCA baselines by
10.94-17.11 % in reconstruction accuracy while reducing communication cost by 42.9% relative to raw state exchange. These findings demonstrate that non-linear embeddings provide an effective foundation for scalable, communication-efficient state dissemination in Cloud-Edge-IoT infrastructures.

Federated learning (FL) enables collaborative model training across distributed healthcare devices without centralizing sensitive data, but remains vulnerable to data poisoning by malicious participants. Existing defenses often rely on a trusted central aggregator, which conflicts with fully peer-to-peer settings. Inspired by swarm intelligence as a pure form of decentralized learning, new approaches can enhance resilience without central control. A trust-based decentralized FL framework is proposed with three layers: peer evaluation, bidirectional sharing enforcement, and gossip-based reputation. In a heart rate prediction task under data poisoning attacks, detection is improved from 33–58% to 92% with gossip, while MAE is reduced beyond the attack-free baseline. These results show that trust and reputation can both defend and enhance learning in fully decentralized FL settings inspired by swarm intelligence.

Distributed edge applications often operate under strict resource and environmental constraints, where reactive orchestration strategies may lead to performance degradation and instability. In this paper, we present a simulation-aware development approach for a distributed edge application, focusing on the orchestration of an urban sound monitoring use case deployed on temperature-constrained and resource-limited devices. To support design-time decision making, the development process is extended with a simulation environment that models system behavior, including workload and CPU temperature, and enables the evaluation of orchestration strategies under realistic conditions. We also introduce a forecast-driven orchestration mechanism that leverages time-series prediction to proactively adapt to changing environmental conditions and workload. The proposed approach is evaluated in both small- and large-scale scenarios. The results show that the forecast-driven strategy maintains low latency and stable operation under increased workload, while reducing unnecessary data migrations and preventing thermal overload by distributing the workload more evenly among devices. These findings demonstrate that integrating simulation into the development process effectively supports the design of reliable and scalable orchestration strategies for distributed edge applications.

The Cognitive Computing Continuum (CCC) promises intelligent orchestration across edge, fog, and cloud tiers, yet its management demands expertise that most infrastructure operators lack. This paper presents an integrated platform, developed within an EU Horizon Europe project, that provides the infrastructure to bridge this skill gap by connecting multi-source telemetry to scenario-based operator training. The platform comprises three layers: (i) the Telemetry Data Collector and Monitoring Engine (TDCME), unifying Prometheus, Kepler energy estimation, and Cilium/Hubble eBPF flows behind a single REST API; (ii) the Dynamic Graph Modeller (DGM), maintaining a live Kubernetes topology graph via Redis-cached edges and Kafka-distributed GEXF export; and (iii) the Virtual Training Environment (VTE), a web-based platform where operators execute end-to-end ML lifecycle workflows with Keycloak SSO. Validation on a dedicated cluster (216 nodes, 245 edges) confirms correct operation across all six inter-layer interfaces. Platform benchmarks show all API endpoints maintaining p95 < 24 ms under 20 concurrent clients, with DGM graph update latency scaling sub-linearly from 58 ms (10 pods) to 190 ms (50 pods). The platform is fully containerized and deployable on standard Kubernetes cluster

The rapid growth of cloud and edge computing has created a fragmented continuum of underutilized and isolated resources. This paper introduces the Association-based continuum, developed within the EMPYREAN EU project, which federates heterogeneous IoT, edge, and cloud resources into secure, autonomous execution environments. Associations, managed by Aggregators, enable dynamic resource sharing, workload placement, and data management across administrative domains. The proposed architecture integrates AI-driven orchestration, decentralized trust, and programmable workflows to support hyper-distributed applications with stringent latency and resilience requirements. By unifying infrastructure providers, service providers, developers, and users, the model enables scalable, trustworthy, and economically viable next-generation distributed computing ecosystems.