Data Spaces: Federated Data Ecosystems

• Data Spaces are decentralized, federated ecosystems that enable controlled data sharing among independent participants while preserving data sovereignty.
• Their architecture integrates connectors, catalogs, discovery services, and policy layers to enforce secure, rule-based, and interoperable interactions.
• Operationally, data spaces support applications from smart-city analytics to secure multi-party computations, balancing performance with trust and compliance.

Data spaces are a polysemous technical term. In the dominant contemporary usage, a data space is a decentralized, federated ecosystem in which multiple, independent participants share and exchange data under common rules while each participant retains control over its own data assets; data remain at source and are shared through secure, policy-governed access mediated by interoperable components rather than by physically consolidating data into a centralized repository (Marojevikj et al., 1 Sep 2025, Fotiou et al., 18 Apr 2025). In adjacent literatures, the same expression denotes feature spaces in which datasets live and are analyzed, topological spaces built over finite datasets, multilevel graph substrates for pay-as-you-go integration, associative “spaces” in pure data foundations, and Riemannian manifold data spaces for statistical modeling (Lombeyda et al., 15 Jan 2026, Chen et al., 2014, Caputo et al., 30 Mar 2025, Youssef, 19 Aug 2025, Fukuzawa et al., 29 Aug 2025). The term therefore names a family of structures whose common concern is organized data interaction, but whose formal objects, guarantees, and intended operations differ substantially.

1. Federated concept and scope

In European and industrial research, a data space is described as a federated digital ecosystem, a sovereign-by-design ecosystem, or a federated, standardized environment that enables organizations to share and jointly utilize data while preserving data sovereignty, multi-level interoperability, and trust (Fotiou et al., 18 Apr 2025, Amaxilatis et al., 29 Nov 2025). This usage is explicitly contrasted with centralized data platforms such as data lakes, which can create new silos and concentrate power, with data meshes, which distribute ownership but generally stop short of cross-organizational usage control and certification, and with data marketplaces, which focus on transactional exchange rather than ongoing, policy-governed collaboration (Amaxilatis et al., 29 Nov 2025). A data space is therefore not a single platform but a federated ecosystem with multiple data spaces that may be sectoral or cross-sector (Ishihara et al., 27 Jan 2025).

The core principles recur across architectures and position papers. Data sovereignty denotes persistent control by the data owner over access and usage; interoperability denotes standard interfaces, shared vocabularies, and common metadata; trust denotes identity, authentication, authorization, auditability, certification, and governance; and decentralization denotes peer-to-peer interaction via connectors rather than compulsory central storage (Martella et al., 20 Mar 2025, Arnold et al., 10 Jul 2025). Complementary formulations add a level playing field, decentralized “soft infrastructure,” and public–private governance as design principles, especially in European strategy documents and Gaia-X-aligned discussions (Marojevikj et al., 1 Sep 2025). This framing explains why initiatives such as the International Data Spaces Association, Gaia-X, FIWARE, the Data Spaces Business Alliance, the Data Spaces Support Centre, iSHARE, XFSC/GXFS, Eclipse Dataspace Components, and Pontus-X are treated as partially convergent rather than mutually exclusive stacks (Martella et al., 20 Mar 2025).

A common misconception is that data spaces are merely a branding layer over ordinary API integration. The literature instead treats them as socio-technical infrastructures in which legal agreements, certification, policy enforcement, and semantic interoperability are co-equal architectural concerns (Ishihara et al., 27 Jan 2025, Marojevikj et al., 1 Sep 2025).

2. Architectural building blocks

Typical data space architectures comprise participants, connectors, catalogs or brokers, discovery services, and policy and trust services (Fotiou et al., 18 Apr 2025). Providers publish offerings, consumers discover them, and connectors mediate contract negotiation, policy enforcement, and transfer coordination between peers; in IDS and EDC-derived architectures, the connector is the secure, policy-enforced endpoint at each participant (Amaxilatis et al., 29 Nov 2025). The control plane manages catalogs, policies, offers, and contract agreements, while the data plane carries application data pipelines, inference outputs, and conventional transfers (Amaxilatis et al., 29 Nov 2025, Arnold et al., 10 Jul 2025).

Several protocol and model families recur. ETSI NGSI-LD provides uniform HTTP-based operations over JSON-LD data items and supports subscriptions to data-driven events, making it central in FIWARE-centric and smart-city deployments (Fotiou et al., 18 Apr 2025). The Dataspace Protocol standardizes catalog discovery, contract negotiation, and transfer process management in EDC-based systems (Amaxilatis et al., 29 Nov 2025). Gaia-X contributes a trust framework and service descriptions, while IDS RAM supplies role decomposition, governance concepts, and connector semantics (Arnold et al., 10 Jul 2025, Martella et al., 20 Mar 2025). These are often combined rather than substituted.

Recent work extends the canonical data-asset focus toward service offerings. In the proposed service abstraction for EDC, a service is a function that consumes input arguments and produces an output, formally $f: (x_1,\dots,x_n)\to y$, and is exposed as a first-class asset governed by the same contract and usage machinery as data offerings (Arnold et al., 10 Jul 2025). Invocation is asynchronous and signal-based, with monotonic state transitions such as INITIALIZING, INITIALIZED, STARTING, RUNNING, FAILED, FINISHED, and CLOSED, allowing provider and consumer connectors to coordinate service execution without collapsing service interaction into ad hoc REST access (Arnold et al., 10 Jul 2025).

At the edge, the same architectural logic is embedded into ETSI MEC. EdgeDS extends the MEC framework with IDS artifacts, introduces an IDS Connector-as-a-Service model managed by the MEC platform, and publishes connector instances as MEC services so that MEC applications can be enriched automatically with discovery, negotiation, and sovereign exchange capabilities (Kalogeropoulos et al., 2023). This suggests that the connector pattern has become the primary architectural invariant across cloud, edge, and hybrid deployments.

3. Sovereignty, trust, and usage control

Usage control is a defining property of data spaces because policy evaluation is not restricted to a one-time permit/deny decision. The literature distinguishes pre-access, in-use, and post-access enforcement, covering purpose limitation, time restriction, access count, rate limit, regional restrictions, deletion, encryption, aggregation, anonymization, activity logging, and delegation of permission as “sticky policies” (Dam et al., 2023). These obligations can be preventive, detective, or continuous, and are typically administered through a Policy Administration Point, evaluated by a Policy Decision Point, enforced by a Policy Enforcement Point, and parameterized by contextual facts from a Policy Information Point (Fotiou et al., 18 Apr 2025, Dam et al., 2023).

A concrete semantics-aware formulation appears in NGSI-LD-based access control. The decision function is

$$Decide(Consumerid, operation, URL) \rightarrow \{true, false\},$$

with authorization propagated by a semantic superset relation across type, object, and attribute URLs, so that permission on a type can imply permission on all objects of that type and their attributes (Fotiou et al., 18 Apr 2025). The associated algorithm is deny-by-default: iterate over the policy set $P$, and return true only if consumer identifier, operation, and URL-superset semantics match; otherwise deny (Fotiou et al., 18 Apr 2025). This model is notable because it couples access control to the semantics of the data model rather than to opaque endpoint paths.

To reduce centralization and preserve owner sovereignty, recent designs decentralize the Policy Administration Point through W3C Verifiable Credentials. In that model, each owner issues credentials to consumers via OID4VCI, consumers present them via OID4VP, and revocation is handled through W3C Status List 2021 bitstrings that the PDP periodically refreshes (Fotiou et al., 18 Apr 2025). The presentation binds the nonce and intermediary URL, which blocks cross-intermediary replay because a malicious intermediary cannot alter the URL without invalidating the signature (Fotiou et al., 18 Apr 2025). This is paired with continuous re-evaluation for subscriptions: if a capability expires or is revoked, the system automatically un-subscribes the consumer from the broker, preserving sovereignty over long-lived event streams (Fotiou et al., 18 Apr 2025).

The policy layer is also expanding semantically. One forward-looking line proposes a SPARQL-based authorization guard over RDF graphs representing assets, contracts, identities, and runtime context, so that permit/deny and obligation derivation become graph queries rather than bespoke middleware logic (Marojevikj et al., 1 Sep 2025). A plausible implication is that semantic policy evaluation could eventually subsume a range of current PIP/PDP attribute-resolution patterns.

Secure computation research pushes sovereignty further by proposing “zero-trust” intermediaries that cannot access users’ data at all. In these architectures, Secure Multi-Party Computation and Fully Homomorphic Encryption are integrated into data spaces alongside VCs, ODRL, OPA/Rego, DPV, and DCAT, so that policy validation, node selection, and access control are cryptographically enforced rather than only contractually asserted (Fabianek et al., 2024).

4. Semantic interoperability, metadata, and discovery

Semantic interoperability is the principal mechanism by which data spaces avoid a single global schema while still supporting coordinated exchange. NGSI-LD binds attributes to globally unique URIs through JSON-LD and is repeatedly used as the semantic substrate for context management in IoT and smart-city deployments (Amaxilatis et al., 29 Nov 2025). DCAT and DCAT-AP structure dataset and catalog metadata; RDF and SPARQL support graph-based integration and query; SHACL validates conformance; IDS Information Model ontologies describe participants, assets, contracts, and usage control; Semantic Hubs and vocabulary repositories curate shared models; and Linked Data Event Streams support near real-time dissemination (Conde et al., 2024, Ishihara et al., 27 Jan 2025, Marojevikj et al., 1 Sep 2025).

Metadata quality is not treated as a secondary documentation problem. The YODA Open Data Portal integrates Apache NiFi, CKAN, DCAT/DCAT-AP, OAI-PMH v2, NGSI-LD, FIWARE Smart Data Models, SHACL-aligned validation, and a metadata quality assessment API to automate generation, publication, and scoring of metadata intended for European Data Spaces (Conde et al., 2024). In November 2023, YODA achieved Findability 100.0, Accessibility 96.0, Interoperability 80.0, Reusability 75.0, and Contextuality 20.0, all above the corresponding EU averages reported in the same study (Conde et al., 2024). The significance is straightforward: discoverability, licensing clarity, contactability, semantic consistency, and machine-readability are prerequisites for any higher-layer contract or policy mechanism.

Semantic storage research extends the same concern into distributed ledgers. A systematic comparison of RDF storage on Ethereum, Hyperledger Fabric, and hybrid Fabric-plus-Ethereum architectures used KBPedia with 1,504,364 triples and showed approximate disk usage of 2.59 GB for public direct storage, 64.20 GB for public smart contracts, 0.93 GB for private storage, and 0.95 GB for the hybrid design (Cano-Benito et al., 3 Jul 2025). Read and reconstruction time likewise favored private and hybrid approaches over direct public logging (Cano-Benito et al., 3 Jul 2025). The paper’s conclusion is specific: private DLTs are the most efficient for storing and managing semantic content, while hybrid DLTs offer a balanced trade-off between public auditability and operational efficiency (Cano-Benito et al., 3 Jul 2025).

Cross-space interoperability adds another semantic and trust dimension. The comparison between Catena-X and Japan’s DATA-EX shows that both sides can share DCAT v2 catalogs, PKI/TLS endpoint assurance, and vocabulary alignment through a cross-index repository, but participant trust diverges because Catena-X uses DID/VC-based SSI and DATA-EX uses OIDC/JWT (Ishihara et al., 27 Jan 2025). The proposed “inter-exchangeable topology” therefore preserves regional trust models instead of forcing immediate homogenization, while relying on private mutual recognition and common metadata structures for practical exchange (Ishihara et al., 27 Jan 2025).

5. Operational extensions and representative deployments

Operationally, data spaces now span service invocation, edge orchestration, smart-city analytics, and cross-domain AI workflows. In the EDC service layer, provider-side adapters can wrap local code or external services and expose them as service assets with argument types, return type, and ODRL-governed usage policies (Arnold et al., 10 Jul 2025). In the prototype MNIST deployment, the mean duration from start until result retrieved was 268.2 ms for EDC-mediated invocation, compared with 25.2 ms for direct HTTP invocation, while the neural network inference time itself was 7.7 ms (Arnold et al., 10 Jul 2025). The reported overhead factor is approximately 10.64, but the architecture is argued to remain acceptable for more complex or long-running services (Arnold et al., 10 Jul 2025).

In smart-city infrastructures, data spaces are used as governance and interoperability layers over the cloud–edge continuum. The smart office implementation built on EDC, NGSI-LD, GAIA-X vocabularies, and AC3 deployed containerized microservices for anomaly detection, forecasting, and occupancy detection near the data source while using cloud orchestration for larger-scale training and lifecycle management (Amaxilatis et al., 29 Nov 2025). Reported anomaly-detection performance was Accuracy approximately 0.90–0.91, Precision up to 0.98, Recall approximately 0.89–0.90, and F1 approximately 0.94–0.95 across specialized models; the real-time presence detector achieved Weighted accuracy 0.72, Precision 0.71, Recall 0.72, and F1 0.71 in the smart office evaluation (Amaxilatis et al., 29 Nov 2025). These results illustrate the paper’s argument that data spaces provide the governance and technical scaffolding for secure, sovereign, and compliant exchange across the cloud–edge continuum (Amaxilatis et al., 29 Nov 2025).

EdgeDS applies the same logic to MEC. In its experiments, preparing IDS services took 9.7 s, catalog configuration took 1.37 s, and IDS-enabled data exchange increased with payload size from 1.788 s at 5 MB to 22.430 s at 150 MB, whereas the baseline direct MEC exchange ranged from 0.088 s at 5 MB to 0.8255 s at 150 MB (Kalogeropoulos et al., 2023). The overhead is explicit, but so is the trade-off: connector preparation, policy enforcement, contract evaluation, and continuous logging buy trust, sovereignty, and certified interoperability that direct MEC services do not provide (Kalogeropoulos et al., 2023).

Secure-computation extensions target scenarios where even connectors or intermediaries should not learn the raw inputs. The proposed three-phase architecture for trustless intermediaries integrates onboarding with Gaia-X-style proofs of membership, setup of datasets, compute nodes and policies, and transaction-time orchestration of MPC or FHE with node selection constraints such as geography, trust zone, and latency (Fabianek et al., 2024). Real-world use cases include air traffic management, manufacturing-as-a-service, and secondary use of data such as health and mobility, where national or organizational boundaries make direct centralized sharing unacceptable (Fabianek et al., 2024).

Across these deployments, a common pattern emerges: data spaces increasingly function as operational substrates for composable services and policy-bearing analytics, not merely as registries of downloadable datasets.

6. Alternative formal meanings and future directions

Outside the federated-infrastructure literature, “data space” acquires markedly different formal meanings. In XR analytics, data spaces are the feature spaces in which datasets live and are analyzed, often high-dimensional manifolds where each observation $x\in\mathbb{R}^D$ is described by many variables; hybrid XR maps these spaces through a projection $f:\mathbb{R}^D\to\mathbb{R}^3$ and couples 3D spatial encodings with embedded 2D dashboards for sensemaking and explainability (Lombeyda et al., 15 Jan 2026). In that setting, the term does not denote inter-organizational exchange infrastructure but an analytic geometry of observations and model outputs.

A more classical formalization treats the data space of a finite dataset $X$ as a finite topological space $D=(X,\mathcal{T})$, where open sets, data functions, and data relations become the substrate for constructing an information space and then a knowledge space $K=K_T\times K_P$ through interpretation, induction, deduction, modeling, and validation (Chen et al., 2014). Another formal line models dataspaces through multilevel graphs equipped with contraction and expansion, allowing local-to-global abstraction, pay-as-you-go integration, and exact traceability from supernodes back to base-level data (Caputo et al., 30 Mar 2025). In the “pure data” foundation, a space is associative data studied through its semiring of endomorphisms, from which natural numbers, integers, rationals, boolean spaces, matrix algebras, Gaussian integers, quaternions, and integer octonions are constructed organically (Youssef, 19 Aug 2025). Statistical learning adds yet another use: Riemannian manifold data spaces are manifolds $(\mathcal{M},g)$ on which coordinate-invariant Rm-NML code lengths are defined with respect to the volume form $dV_g$ rather than Lebesgue measure (Fukuzawa et al., 29 Aug 2025).

These usages are formally distinct from the federated-ecosystem meaning, and conflating them obscures the literature. The shared thread is structural: each tradition defines a space in which data, relations, transformations, or policies can be organized without reducing the problem to an undifferentiated repository.

Future work in the dominant federated sense is converging around several fronts. Standards convergence remains central, with DSBA and DSSC attempting to harmonize Gaia-X, IDS, FIWARE, and related building blocks (Martella et al., 20 Mar 2025). Semantic data models and knowledge graphs are repeatedly proposed as the next increment in interoperability, alongside SPARQL-based authorization for finer-grained control over linked policies and assets (Marojevikj et al., 1 Sep 2025). Collaboration, provenance, and storytelling inside XR workspaces remain open in analytic data-space research (Lombeyda et al., 15 Jan 2026). At the same time, cross-space interoperability still depends on assurance mapping, policy translation, and vocabulary alignment rather than on a universally accepted global stack (Ishihara et al., 27 Jan 2025). This suggests that the future of data spaces will likely remain plural: increasingly interoperable, but not monolithic.