bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 3쪽 중 3쪽
, activities, and agents. Use classes like prov:Agent,
prov:Person, and prov:Organization to specify roles. For example, prov:wasAttributedTo can
link an AI model to its developer [20].
• FOAF (Friend of a Friend): An RDF vocabulary for describing people, their activities, and their relations to
other people and objects. Could use foaf:Person and foaf:Organization to represent individuals and
organizations. Properties like foaf:maker can denote the creator of an AI agent [6].
• Schema.org: Provides a collection of schemas for structured data on the internet, covering entities,
relationships, and actions. Could use types such as Person, Organization, and SoftwareApplication.
Properties like schema:creator and schema:maintainer can define roles related to AI agents [22].
• IEEE P7007 Ontology for AI Agent Governance: A standard being developed by IEEE for ontology-based
governance of AI systems [11].
This is the list of entities and activities defined by W3C PROV-O:
14
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
Name Can be used instead of the role
WasGeneratedBy Developer
Used
WasInformedBy Trainer
WasDerivedFrom Model
WasAttributedTo Owner
WasAssociatedWith Generic relationship
ActedOnBehalfOf User
Expressing relationships can provide higher flexibility than expressing fixed roles. This would give us a different
interface (the front-end could simply report the predefined relationships only or provide a full view of all relationships).
4 Conclusion
The proliferation of AI agents across diverse applications demands a comprehensive trust framework to ensure
accountability, privacy, and interoperability. This paper has focused on defining the foundational role of identity within
such a framework, emphasizing the importance of verifiable identifiers as anchors for reputation and trust systems.
These identifiers enable agents to participate in decentralized ecosystems where trust is established through attestations
and verifiable relationships with other entities, including humans, organizations, and other agents.
Through a comparative analysis, we evaluated existing solutions for AI agent identity. Each approach has strengths
and limitations, particularly when addressing issues such as privacy, key security, and attribute expressiveness. Our
proposed hybrid model combines DIDs with public on-chain attestations, enhanced by privacy-preserving tools like
Verifiable Credentials (VCs) and Zero-Knowledge Proofs (ZKPs). This framework balances transparency and privacy,
enabling flexible, scalable, and compliant identity solutions.
Beyond the proposed architecture, there remain several critical areas for future development to evolve this trust
ecosystem. Advancements in Zero-Knowledge Machine Learning (zkML) offer promising opportunities for more
robust behavioral identity mechanisms, enabling attestations that link agent behavior to its verifiable identity without
compromising sensitive information. At the same time, establishing clearer rules and standards for the trust ecosystem,
including semantic definitions of relationships, trust registries, and governance frameworks, will be essential for
fostering a cohesive and interoperable ecosystem.
Further, the governance of the trust ecosystem itself demands thoughtful consideration. Trust registries that authenticate
and validate attestations must operate transparently, while balancing decentralization with mechanisms to ensure
credibility and prevent abuse. Governance models, potentially leveraging decentralized autonomous organizations
(DAOs) or similar constructs, could help standardize rules while allowing flexibility for domain-specific needs.
This paper has provided a foundational architecture for establishing AI agent identity and trust in decentralized systems.
However, the evolution of this ecosystem will require collaboration across disciplines, technological advancements,
15
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
and the development of shared principles and governance models. By addressing these challenges, the community can
create a resilient, adaptable trust framework capable of supporting the next generation of AI-driven systems in complex,
global ecosystems.
Acknowledgments
We are deeply grateful to the reviewers Evin McMullen and Brian Seong for their thoughtful and rigorous feedback
throughout the development of this technical report.
16
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
Annex A: Design Principles and Constraints
The proposed solution was designed to maximize these goals and in accordance to these principles:
Goal Constraints and Challenges Design Principles
Minimize trust assumptions Developers can’t be trusted Either the authentication factor is unknown
by the developer OR it’s based on what
the agent IS and not on what the agent
KNOWS.
Agents can lie The authentication factor to the identifier
is not part of the training set nor the
agent inference model (i.e., the agent is
impersonating another agent).
Reach mass adoption Resistance to change in developers We can’t assume or require the use
of hyper-specialized components.
Well-consolidated, commoditized tech is
preferred over edge-tech and high-end
components.
Complexity of standardization We should focus on lean standards
that focus on the minimum necessary
agreements and allow for multiple
implementation strategies.
Minimum requirements for
verification
Embedded systems, heterogeneous UX Minimize the assumptions about the user
experience, since these agents can be deeply
embedded in different applications and
devices.
Real world trust
digitalization
Complex trust ecosystems, no single root
of trust
Decentralized identity models that allow for
open ecosystems of trust.
No clear attestation frameworks nor
reputation models
Identify the minimum valuable claims and
reach standardization through adoption.
17
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
Annex B: Proposed Architecture based on Privado ID
A Universal ZK Attestation Registry
The Universal ZK Attestation Registry is the central component for managing and verifying identity-related data
in a decentralized ecosystem. It consists of a Primary Registry, which serves as the authoritative source for queries,
attestations, and state information.
The ZK Attestation Registry includes a Registry Explorer (similar to other blockchain block explorers) that allows to
query the registry through UI and API endpoints, so users and applications don’t have to deal with the complexities of
direct blockchain queries.
B Identity Wallets
Identity Wallets function as secure storage mechanisms for verifiable credentials issued to humans or AI agents.
These credentials are managed off-chain, allowing users or agents to maintain privacy while proving specific claims as
needed. Using cryptographic tools, such as Zero-Knowledge Proofs (ZKPs), Identity Wallets enable users to provide
verifications (e.g., “I hold a valid certification”) without exposing sensitive underlying data.
The wallets support key features, including DID-based authentication, credential issuance and storage, and dynamic
credential updates or revocations. They interface with the Universal On-chain Registry to facilitate on-chain verifications
while maintaining the confidentiality of private information. By bridging private credential storage with the registry’s
immutable attestations, Identity Wallets ensure compliance with privacy and data protection requirements.
C Verifier Toolset
The Verifier Toolsetincludes the necessary components for verifying identities and credentials. It provides functions
for DID authentication, ZKP validation, and query formulation through the ZK Query Language. This toolset
interacts with the Universal On-chain Registry to retrieve public attestations and validate private claims presented by
Identity Wallets.
By supporting both on-chain and off-chain verification processes, the Verifier Toolset ensures the accuracy and
compliance of identity verifications across various applications.
18
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
D Issuer Toolset
The Issuer Toolset is designed for credential issuers to create, issue, and manage verifiable credentials. Issuers can
define schemas for credentials, manage credential lifecycles, and record attestations on the Universal On-chain Registry.
These attestations serve as immutable proof of a credential’s existence, while sensitive data remains off-chain.
The toolset also interacts with Trust Registries to ensure that only authorized entities can issue credentials, preserving
the reliability and integrity of the identity system.
D.0.1 Trust Registries
Trust Registries are on-chain directories that manage the list of authorized participants within a given ecosystem,
such as credential issuers, verifiers, and wallet providers. These registries are responsible for validating the roles and
permissions of participants, ensuring the integrity of the credential issuance and verification processes. Each Trust
Registry operates as part of a larger Ecosystem of Trust, where governance models can vary depending on the specific
application.
The governance of a Trust Registry is determined by the stakeholders within the ecosystem it serves. For instance,
in a regulatory-compliant financial system, governance may be centralized under financial authorities. Alternatively,
a decentralized reputation system may use a DAO model to vote on updates and rule changes. The ability to define
governance rules at the ecosystem level allows Trust Registries to align with the operational requirements and trust
standards of their specific context.
19
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
References
[1] A Comprehensive Overview of Large Language Models. Retrieved from https://arxiv.org/html/2307.
06435v7. Accessed on December 29, 2023. Dec. 2023.
[2] Art. 4 GDPR – Definitions - General Data Protection Regulation (GDPR) . Retrieved from https://gdpr-
info.eu/art-4-gdpr. Accessed on March 29, 2018. Mar. 2018.
[3] Mark Coeckelbergh. “Artificial Intelligence, Responsibility Attribution, and a Relational Justification of
Explainability.” In: Science and Engineering Ethics 26.4 (2020), pp. 2051–2068. DOI: 10. 1007/ s11948-
019-00146-8.
[4] Finale Doshi-Velez et al. “Accountability of AI Under the Law: The Role of Explanation.” In:arXiv preprint
(2017). arXiv: 1711.01134 [cs.AI]. URL: https://arxiv.org/abs/1711.01134.
[5] Encyclopedia Editorial Office. Identity. Retrieved from https : / / encyclopedia . pub / entry / 54087.
Accessed on January 25, 2024. Jan. 2024.
[6] FOAF Vocabulary Specification. Retrieved from http://xmlns.com/foaf/spec. Accessed on December 22,
2023. Dec. 2023.
[7] Fund MPC Wallets | Coinbase Developer Documentation. Retrieved from https://docs.cdp.coinbase.
com/mpc-wallet/docs/fund-mpc-wallets . Accessed on September 20, 2024. Sept. 2024.
[8] David J. Gagne et al. “Machine Learning for Stochastic Parameterization: Generative Adversarial Networks
in the Lorenz ’96 Model.” In: Journal of Advances in Modeling Earth Systems 12.3 (2020). DOI: 10.1029/
2019ms001896.
[9] Nicole Gillespie. A Review of Trust in Artificial Intelligence: Challenges, Vulnerabilities and Future Directions.
Retrieved from https : / / www . academia . edu / 64416954 / A _ Review _ of _ Trust _ in _ Artificial _
Intelligence_Challenges_Vulnerabilities_and_Future_Directions. Accessed on January 1, 2021.
Jan. 2021.
[10] T. Hauer. “Importance and Limitations of AI Ethics in Contemporary Society.” In: Humanities and Social
Sciences Communications 9 (2022). DOI: 10.1057/s41599-022-01200-7 .
[11] IEEE Standards Association. IEEE Ontological Standard for Ethically Driven Robotics and Automation Systems.
Retrieved from https://standards.ieee.org/ieee/7007/7070 . Accessed on December 3, 2024. Dec.
2024.
[12] S. Jain et al. “AI and Democracy’s Digital Identity Crisis.” In: arXiv (2023). arXiv: 2311.16115 [cs.CY]. URL:
https://arxiv.org/abs/2311.16115v1.
[13] Lab, Gaudiy Web3 and AI. Potential and Challenges of zkML - Gaudiy Web3 and AI Lab - Medium. Retrieved
from https://medium.com/gaudiy-web3-and-ai-lab/88022a0033a5 . Accessed on 2024. 2024.
[14] Will Ladd. Introducing Zero-Knowledge Proofs for Private Web Attestation with Cross/Multi-Vendor Hardware.
Retrieved from https : / / blog . cloudflare . com / introducing - zero - knowledge - proofs - for -
private-web-attestation-with-cross-multi-vendor-hardware/ . Aug. 2021.
[15] Derek Leben. The Ethical Challenges of AI Agents . Tepperspectives. Retrieved from https : / /
tepperspectives.cmu.edu/all-articles/the-ethical-challenges-of-ai-agents/ . Feb. 2025.
[16] T. Lit. Key Management and Identity Strategies for Crypto Agents . Spark by Lit Protocol. Retrieved from
https://spark.litprotocol.com/agent-identity. Accessed on 2024. 2024.
[17] Parikshit N. Mahalle, Gurunath Shinde, and Parikha M. Shafi. “Rethinking Decentralised Identifiers and Verifiable
Credentials for the Internet of Things.” In: Internet of Things, Smart Computing and Technology: A Roadmap
Ahead. Springer, 2020, pp. 317–334. DOI: 10.1007/978-3-030-39047-1_16 .
[18] Claudio Mazzocca et al. “A Survey on Decentralized Identifiers and Verifiable Credentials.” In:arXiv (2024).
arXiv: 2402.02455 [cs.CR]. URL: https://arxiv.org/abs/2402.02455v1.
[19] Dinesh Naicker and Mikhail Moodley. “Challenges of User Data Privacy in Self-Sovereign Identity Verifiable
Credentials for Autonomous Building Access During the COVID-19 Pandemic.” In: Frontiers in Blockchain
7 (2024). URL: https://www.frontiersin.org/journals/blockchain/articles/10.3389/fbloc.
2024.1374655/full.
[20] PROV-DM: The PROV Data Model. Retrieved from https : / / www . w3 . org / TR / 2013 / REC - prov - dm -
20130430. Accessed on May 10, 2024. May 2024.
[21] Khaled Salah et al. “Blockchain for AI: Review and Open Research Challenges.” In: IEEE Access 8 (2020),
pp. 10127–10149. DOI: 10 . 1109 / ACCESS . 2020 . 2966174. URL: https : / / ieeexplore . ieee . org /
document/8965133.
[22] Schema.org Community Group. Schema.org. Retrieved from https://schema.org. Accessed on December 3,
2024. Dec. 2024.
20
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
[23] T. South et al. “Verifiable evaluations of machine learning models using zkSNARKs.” In:arXiv (2024). arXiv:
2402.02675 [cs.AI]. URL: https://arxiv.org/abs/2402.02675v2.
[24] Soumya Y . Tadimalla and Mary Lou Maher. “AI and Identity.” In:arXiv (2024). arXiv: 2403.07924 [cs.AI].
URL: https://arxiv.org/abs/2403.07924v2.
[25] Tae-Won Um et al.Trust Management for Artificial Intelligence: A Standardization Perspective. Retrieved from
https://www.academia.edu/87233400/Trust_Management_for_Artificial_Intelligence_A_
Standardization_Perspective?nav_from=e62bc347-8e40-4a36-b112-8423b81e6e14 . Accessed on
December 3, 2024. Dec. 2024.
[26] Shannon Vallor and Till Vierkant. “Find the Gap: AI, Responsible Agency and Vulnerability.” In: Minds &
Machines 34.3 (2024), pp. 1–23. DOI: 10.1007/s11023-024-09674-0 .
[27] Yutao Xing, Yan Wu, and Yuchen Liu. “Zero-Knowledge Proofs in Machine Learning: A Survey.” In:Journal of
Artificial Intelligence Research 67 (2023), pp. 1–24.
[28] ZKProof Standards. ZKProof Standards: Scaling Trustless DNN Inference, zkml applications at ZKProof.org
by Daniel Kang - ZKProof Standards. Retrieved from https://zkproof.org/2023/09/18/zkml-where-
are-we-now-where-do-we-go-from-here-talk-summary-zkproof-5-5-daniel-kang . Accessed on
September 18, 2023. 2023.
21프로젝트로 돌아가기