Identity Ecosystem Graph Analysis

• Identity Ecosystem Graph is a graph-based model that represents digital identities, attributes, credentials, and disclosure relationships across complex sociotechnical systems.
• It employs formal graph-theoretic methods with weighted, directed edges to quantify disclosure probabilities and assess privacy risks.
• Applications span digital identity federation, SSI ecosystems, blockchain identity inference, and cross-platform linkage, highlighting its role in modern IAM research.

Searching arXiv for the provided and closely related papers on identity ecosystem graphs, digital identity ecosystems, SSI ecosystems, IAM graphs, PKGs, trust graphs, and blockchain identity inference. An identity ecosystem graph is a graph-based representation of identities, identity-bearing entities, attributes, credentials, trust assertions, and their consequences across a sociotechnical environment. In the most explicit formulation, it is a directed, weighted graph whose nodes are personally identifiable information (PII) attributes and whose edges encode empirical disclosure relationships learned from identity theft and fraud cases (Niu et al., 6 Aug 2025). In adjacent literatures, closely related structures appear as AI identity ecosystems, digital identity zone models, self-sovereign identity ecosystems, personal knowledge graph ecosystems, IAM interaction graphs, blockchain account interaction graphs, trust graphs, and zero-infrastructure capability graphs. Taken together, these models treat identity not as a single identifier but as a layered relational object distributed across actors, artifacts, infrastructures, policies, and impacts (Tadimalla et al., 2024, Kohli, 2011, Laatikainen et al., 2021, Skjæveland et al., 2023, Madireddy, 11 Dec 2025, Zhang et al., 2019, Kuri, 5 Jan 2026).

1. Formal definitions and graph-theoretic core

A recurring formal core is the graph $G=(V,E)$. In the privacy-risk setting, $V$ is the set of PII attributes and $E$ is the set of directed disclosure relations; an edge $A \rightarrow B$ with weight $w_i$ means that an event disclosing $A$ may lead to a disclosure of $B$, and such a disclosure happened $w_i$ times in the empirical data. The same paper defines conditional disclosure probabilities by

$$P(B_j \text{ exposed} \mid A \text{ exposed}) = \frac{w_j}{\sum_{l=1}^k w_l},$$

making the graph simultaneously structural and probabilistic (Niu et al., 6 Aug 2025).

More general identity ecosystem formulations preserve the same idea while broadening the node set. In the AI identity ecosystem, the global node set is

$V = C \cup I \cup M \cup F \cup Q \cup H \cup G,$

where creators, infrastructure, models, frameworks, consequences, human identity groups, and governance nodes are linked by directed relations such as build, enable, constrain, impact, affect, feedback, and regulation. In self-sovereign identity, the core interaction graph is the issuer–holder–verifier triangle governed by a governance authority. In zero-infrastructure capability graphs, vertices are attestation or vouch statements and edges are induced by vouches referencing other statements. This suggests a common abstraction: identity ecosystems are graphs in which identity emerges from relations among entities, claims, and authorities rather than from isolated labels (Tadimalla et al., 2024, Laatikainen et al., 2021, Kuri, 5 Jan 2026).

The literature also distinguishes graph structure from graph instances. The vehicle-centric model defines a schema graph $V$0 over entity types and edge types, and an instance graph $V$1 over concrete actors and flows. The same schema/instance separation is implicit in ontology-based innovation ecosystems and PKG ecosystems, where classes such as Organization, Talent, Patent, Article, and Project define the type system, while concrete organizations, persons, and artifacts instantiate it (Yuan et al., 2024, Tejero et al., 2020, Skjæveland et al., 2023).

2. Structural motifs: zones, layers, owners, and global identities

One major motif is partitioning. The Digital Identity Zone Model partitions the digital world into Friends/Family / Social Zone, Purchase / Consumer / Telecom Zone, Corporate Zone, Services / Banking / Certification Zone, and Government Zone, each with characteristic data classifications such as Public, Private, Sensitive, and Confidential. Within this model, the paper states the identity equation “User = Many Identities = Many Roles = Many Resources = Many Access Mechanisms,” and formalizes the accumulation of user IDs, passwords, and roles across zones. Graphically, a single global identity node $V$2 links to multiple credentials and roles distributed across zone-specific subgraphs (Kohli, 2011).

A second motif is layering. SSI ecosystems are described as innovation ecosystems organized around four building blocks—Conceptualization, Technology & Governance Dual Framework, Business Models, and Collaborative Ecosystem—and aligned with the TrustOverIP four-layer architecture in which the bottom two layers provide technical trust and the upper layers represent human trust. PKG ecosystems use a comparable layered view: the PKG and its management system sit between a population layer of private and public data sources and a utilization layer of personalized services, with the owner retaining full read and write access and the exclusive right to grant others read and write access to specified parts of the graph (Laatikainen et al., 2021, Skjæveland et al., 2023).

A third motif is the distinction between internal and external identity. The AI identity literature defines AI identity internally as the collective characteristics, values, and ethical considerations embodied in the creation of AI technologies, and externally as shaped by individual perception, societal impact, and cultural norms. It then organizes the ecosystem through the triplet Creators, Creations, and Consequences. This yields a layered graph in which identity flows from creators to infrastructure, models, and frameworks, and then to ethical, social, cultural, economic, and psychological consequences (Tadimalla et al., 2024).

These motifs are complementary rather than exclusive. Partitioning emphasizes contexts of use, layering emphasizes technical and governance strata, ownership emphasizes administrative control, and internal/external identity emphasizes how identity is produced and perceived. Across the literature, an identity ecosystem graph is therefore less a single canonical data structure than a family of structurally related graph models.

3. Trust, delegation, governance, and revocation

Identity ecosystem graphs are not only descriptive; they are normative and operational. In digital identity federation, OIX introduces a governance layer with Central Authentication Bodies, OITF Providers, assessors, and auditors. Trust is represented through authentication edges from user to IdP, federation edges from IdP to relying party, role edges to services, and policy edges from governance bodies to IdPs and relying parties. Levels of assurance become edge or node attributes, so admissible access paths can be constrained by assurance thresholds (Kohli, 2011).

SSI ecosystems make this governance explicit. The governance authority publishes a governance framework containing business, legal, and technical policies for issuing, holding, and verifying credentials. The trust triangle—Issuer, Holder, Verifier—sits inside this governance framework, while DIDs, Verifiable Credentials, DIDComm, wallets, agents, and a trusted registry implement the technical substrate. The literature is explicit that technology and governance must be defined “side by side,” which corrects the common misconception that decentralization eliminates governance requirements (Laatikainen et al., 2021).

The Mediterraneus protocol extends this pattern into an SSI-native service economy. A user controls both an EOA and an SSI identity with a DID and identity key pair. Identity Smart Contracts maintain mappings $V$3 and $V$4, while on-chain service publication and purchase require hasValidStatus(EOA). Off-chain access is mediated by a connector that verifies a Verifiable Presentation, checks revocation status via the Identity Smart Contract, verifies wallet control through a challenge signature, and then confirms proof of purchase through an ERC-20 access token balance. Identity, economic access, and service consumption are thus joined in a single graph of DIDs, VCs, VPs, smart contracts, wallets, services, and connectors (Giorgino et al., 2024).

Vouchsafe pushes the same idea into disconnected environments. Its Zero-Infrastructure Capability Graph is a DAG of self-verifying signed statements. Let $V$5 be a set of tokens and $V$6 the subset that survives revocation and burn processing. Vertices are tokens of kind attest or vouch, while edges are induced when a surviving vouch token references another surviving token. Acceptance is purely local: $V$7 and depends only on the presented token set, the verifier’s trust roots, and the requested scope. Scope monotonically decreases along delegation paths, revocation removes specific statements, and burn removes all statements issued by an identity (Kuri, 5 Jan 2026).

4. Learning and inference over identity ecosystem graphs

Recent work increasingly treats identity ecosystem graphs as learnable objects. In cloud IAM analytics, logs are modeled as a dynamic heterogeneous graph

$V$8

where nodes include users, roles, and resources, and edges encode access actions over time. A GAT-style message-passing layer updates node embeddings by

$V$9

with attention coefficients $E$0 computed from neighboring embeddings. The framework reports Precision $E$1, Recall $E$2, F1-Score $E$3, and False Positive Rate $E$4, outperforming Random Forest, XGBoost, and LSTM baselines on the reported benchmark. Its ablation study finds that attention adds $E$5 recall and adaptive retraining cuts false positives by $E$6 (Madireddy, 11 Dec 2025).

Blockchain identity inference uses related but domain-specific graphs. BlockGC builds an account interaction graph over EOAs, contract accounts, transaction edges, and contract-call edges, then samples local subgraphs and trains with a joint objective

$E$7

combining supervised account classification and subgraph contrastive learning. Identity here means account type—Exchange, ICO-wallets, Mining, or Phish-hack—rather than a real-world name. DBG4ETH generalizes this into a double-graph model with a Global Static Graph $E$8 and a Local Dynamic Graph $E$9, then calibrates the outputs of both branches before final classification. It reports $A \rightarrow B$0-score improvements of $A \rightarrow B$1 to $A \rightarrow B$2 over single-graph processing and $A \rightarrow B$3 to $A \rightarrow B$4 over similar account identity inference methods (Zhou et al., 2021, Miao et al., 2024).

Cross-platform user identity linkage offers another variant. GraphUIL models each online social network as a graph $A \rightarrow B$5 or $A \rightarrow B$6, learns node representations with a Multi-Stage Aggregation layer that combines global-topology-aware and local-topology-aware aggregation, and then learns a mapping between embedding spaces using known anchor links. On the reported Instagram–Twitter task, GraphUIL attains Accuracy $A \rightarrow B$7 and F1-score $A \rightarrow B$8, outperforming PALE, FRUIP, Node2Vec, and ablated variants (Zhang et al., 2019).

At the representation level, RFA separates identity and position embeddings spectrally. The paper argues that low-frequency information characterizes node positions and high-frequency information characterizes node identities, and proposes training-free Random Feature Aggregation with high-pass and low-pass filters. RFA(H) derives identity embeddings, while RFA(L) derives position embeddings, both with one feed-forward propagation and no learnable parameters (Qin et al., 27 May 2025).

5. Privacy, bias, and adversarial structure

The most direct privacy-risk identity ecosystem graph is constructed from over 5,000 empirical identity theft and fraud cases; the largest reported graph is built from 5,636 cases and has $A \rightarrow B$9 and $w_i$0, while a loss-filtered graph with losses greater than \$10,000 has $w_i$1 and $w_i$2. Link prediction over this graph is performed with featureMLP, featureGCN, and SeeGCN; on the largest graph, the reported validation accuracies are $w_i$3, $w_i$4, and $w_i$5, respectively (Niu et al., 6 Aug 2025).

The same path-based logic appears in vehicle-centric data sharing. There the graph includes Vehicle, Vehicle Component, Additional Vehicle Sensor, Traffic Monitoring Sensor, Charging Facility, Communication Infrastructure, Digital Asset, Person, Organisation, and Data Package nodes, with directed edges for collection, transfer, and sharing. The model is designed specifically to support path and topological analysis of privacy risks, such as identifying which organisations can obtain a person’s data through multi-hop propagation (Yuan et al., 2024).

In AI identity research, harms are not evenly distributed. The identity space includes race, gender, class, sexuality, disability, nationality, and age, and the paper emphasizes that the impact of bias in the creator’s layers “snowballs exponentially” into the consequences layer. It also describes the amplification metaphorically as “bio-accumulation” or “bio-magnification.” This makes the identity ecosystem graph not only a technical dependency graph but also a harm-distribution graph across intersecting social identities (Tadimalla et al., 2024).

Adversarial structure is formalized most sharply in Sybil-resilient trust graphs. A trust graph $w_i$6 models identities and mutual trust; a community $w_i$7 grows over time while trying to keep byzantine penetration

$w_i$8

below a chosen threshold. The paper gives two sufficient tools: graph conductance and vertex expansion. It emphasizes keeping the fraction of byzantines below one third, since that would allow Byzantine Agreement and sybil-resilient social choice. Here the identity ecosystem graph is an admission-control object: its topology is the mechanism by which fake or duplicate identities are contained (Zhang et al., 2019).

6. Misconceptions, limitations, and research directions

Several misconceptions recur across the literature. One is to treat identity as a single central profile. The PKG literature explicitly rejects the earlier “spiderweb” view in which every fact must connect directly to the user; what makes a graph personal is administrative control, not graph topology. Another is to equate decentralization with absence of governance. SSI ecosystems insist on a technology and governance dual framework, and Web3-oriented systems such as Mediterraneus still rely on issuers, governance choices, and status checks. A further misconception is to assume that user control automatically eliminates concentration: SSI research explicitly identifies the “War of Wallets” as a possible outcome in which large technology firms consolidate power at the wallet layer (Skjæveland et al., 2023, Laatikainen et al., 2021, Giorgino et al., 2024).

Limitations are likewise recurrent. IAM graph detection depends on log quality and completeness, assumes a relatively stable entity schema, and incurs higher computational overhead than simple heuristics. The PII disclosure graph is a static snapshot over accumulated cases and does not explicitly model temporal change. Vouchsafe achieves offline verifiability, but does not prescribe distribution, freshness guarantees, or principal continuity after burn. Innovation and vehicle-centric ecosystem graphs face incomplete and confidential data, while PKG ecosystems must reason over inconsistent and partially available personal graphs (Madireddy, 11 Dec 2025, Niu et al., 6 Aug 2025, Kuri, 5 Jan 2026, Tejero et al., 2020, Yuan et al., 2024, Skjæveland et al., 2023).

A consistent research direction is therefore toward richer but still analyzable graphs. The literature points to dynamic and context-aware graphs, federated learning and privacy-preserving graph aggregation, explainable graph models, multi-issuer governance, richer scope and policy semantics, fault-tolerant reasoning over inconsistent subgraphs, process mining over time-stamped identity events, and comparative tools for ecosystem analysis. This suggests that the identity ecosystem graph is evolving from a descriptive metaphor into a general analytic substrate for privacy risk estimation, credential governance, zero-trust monitoring, decentralized service access, and identity-centered knowledge management (Niu et al., 6 Aug 2025, Madireddy, 11 Dec 2025, Laatikainen et al., 2021, Giorgino et al., 2024, Yuan et al., 2024, Tejero et al., 2020).