Capabilities for Distributed Authorization

3.1. IoT On-boarding

On-boarding IoT devices into an overall system requires authentication and authorization; this may need to be mutual.¶

In such scenarios, new devices rarely have connectivity before completing the on-boarding process. It follows that authentication and authorization must work in a fully offline fashion, which in turn requires that authorization tokens provided to the device contain all information required for the authorization step.¶

This specific problem is also addressed in [POA-IOT] and related work.¶

3.2. UAV Control Handover

A similar argument applies to control handover of unmanned aerial vehicles (UAV). The concept of Beyond Visual Line of Sight (BVLOS) missions is to send drones into places that are harder or more costly to reach for humans.¶

Control handover refers to transferring operational control for a drone from one ground control station to (GCS) another. Control handover bears similarities to IoT on-boarding in that the drone is on-boarded to a new control system (and the previous system relinquishes control).¶

In general, aviation authorities such as FAA, EASA, etc. expect control handover to occur under ideal circumstances, in which centralized authorization schemes suffice because connectivity is guaranteed. There are, however, classes of scenarios where this cannot be expected.¶

3.3. Zero Round-Trip-Time (0-RTT)

If fast authorization is a goal, reducing the number of roundtrips to establish a privilege follows. Reducing this to 0-RTT implies being able to store and send verifiable authorization data, rather than engaging in a more complex handshake.¶

Of course, authorization can only follow when authentication already occurred. Authentication in a 0-RTT protocol is predicated on prior key exchange and verification.¶

Both [WIREGUARD] and DTLS 1.3 [RFC9147] offer 0-RTT handshakes. In the former, keys are pre-shared out of band, because WireGuard is used to establish static VPN tunnels. Because mutual authentication is assumed to be part of this process, authenticated encryption is sufficient to ensure that the keys are safely associated with network addresses in a 0-RTT roundtrip.¶

By contrast, DTLS simply offers different kinds of handshakes. 0-RTT can only be used for reconnection when a previous full handshake has provided sufficient authentication.¶

In either case, when 0-RTT authentication is offered, 0-RTT authorization should also be offered in order not to introduce more latency at this latter stage.¶

3.4. Bike Rental

Standing in for similar schemes, urban bike rental is a candidate for distributed authorization.¶

As a user, you typically use a mobile application to purchase some usage time via an internet enabled service. You then use the application to open a lock on a candidate bike, use it, and lock it again on return.¶

But the authorization to use bikes from the service does not expire with the return of a bike. It is a feasible and supported use case to lock one bike e.g. to go into a shop, then find a new bike to return the shopping home, all within the same booked time slot.¶

If centralized authorization systems are used, each of these unlocking operations will require internet connectivity. If connectivity is not given, access to the service will be denied, even though contractually speaking, it must be given.¶

3.5. Principle of Least Authority (POLA)

The Principle of Least Authority refers to a design method for authorization systems in which a process is given only the authority required to perform its purpose, and no more.¶

This is in contrast to identity-based authorization systems, in which processes can perform any activity that an identified user is permitted to perform. More advanced systems such as Role-Based Access Control (RBAC) function similar to identity-based authorization, except that each identified user can hold multiple roles, and roles are typically more strictly limited.¶

On a spectrum where identity-based authorization exists on one end, and RBAC occupies some middle ground, POLA exists on the opposite end of identity-based authorization.¶

Capabilities are eminently suitable for use with POLA, because they can encode arbitrarily fine-grained access control data.¶

In principle, POLA is very similar to how modern mobile operating systems encode permissions: each app must explicitly request e.g. access to the camera, internet, file system, etc. However, where such permissions are applied only locally, capabilities can be transmitted to be applied on remote nodes. In effect, a local application can be granted access to a networked, rather than merely a local camera.¶

3.6. Human Rights Considerations

[RFC8280], updated by [I-D.draft-irtf-hrpc-guidelines-20], lists a number of distinct objectives that help support human rights in protocol design. The above distributed authorization scheme addresses a number of them, such as Connectivity, Reliability, Content agnosticism, Integrity, Authenticity, Pseudonymity, Censorship Resistance, Outcome Transparency, Adaptability, Decentralization and Security, and by way of producing this document, Open Standards.¶

Rather than address each in detail, suffice to say that the use of pseudonymous public keys, and proofs based on cryptographic signatures, the majority of these objectives are reached.¶

It remains to highlight that the scheme outlined in this document observes the end-to-end principle, precisely by temporally decoupling different concerns. This permits for almost arbitrarily disrupted connectivity, and thus also censorship resistance. As capabilities can travel entirely out-of-band to any resource data, e.g. by sneakernet or similar means, they can be a building block of protocols that provide better human rights protections than systems that rely on temporal coupling of authorization concerns.¶

4.1. Authentication

Password-based authentication mechanisms require that the tuple of user name and password (or password hash) are sent to some central repository where records of such tuples are kept; if the tuple is found, the user name is authenticated.¶

This common scheme mixes different aspects to authentication, however, which are worth disambiguating.¶

Identification:

The act of establishing the attributes that describe a person in sufficient detail to identify them.¶

Endowment:

The act of associating attributes derived during the identification phase with an identifier such as may be used in secret proving.¶

Secret Proving:

Logging in to a system requires proving that one is in possession of some secret. It usually also requires providing some identifier that the secret is associated with -- this permits tying secret proving back to identification and endowment steps.¶

This distinction becomes somewhat more relevant when we move towards distributed authentication schemes, which rely on public key cryptography. It's worth noting that it is possible to construct systems that rely solely on secret proving, even without identifiers of any sort.¶

What is typically understood as "authentication", however, is a combination of secret proving and endowment. Specifically, once the secret is proven, a user database is consulted to endow the identifier with associated attributes (i.e. a user profile).¶

In particular, identification occurs in a preparation phase, according to the system's policies -- sometimes prior to when an identifier is first associated with a secret, sometimes after that event, and often before access to the system's functions is granted.¶

Note that a finer grained model is possible in which we distinguish between when identifiers are endowed with attributes in principle vs. for use of the system, but that is unlikely to clarify this document. We prefer a simpler model here.¶

4.2. Authorization

Authorization occurs only after secret proving. Once an identity has been established, it is then mapped to associated privileges, which determine which object(s) it has access to.¶

There exist a plethora of methods to establish this mapping. Access-control lists (ACL) simply provide tuples of identities, privileges and associated objects. Role-based access control (RBAC) is effectively identical, if the identities specified are not those of individuals, but of groups (as a group member, an individual inhabits the associated role). A comparable approach is Organization-based access control (OrBAC), which not only abstracts the identity to that of a role, but performs a similar abstraction on the object and privilege.¶

More complex systems such as context- or lattice-based access control (CBAC and LBAC respectively) derive a mapping from properties of or labels attached to the individuals and objects. Finally, graph-based access control (GBAC) starts with a graph of an organization, and derives privileges from the properties inherited by being part of a larger organizational structure.¶

What these systems address is the problem of managing the mapping of an identity to access privileges for objects, where each system has advantages and disadvantages for various use cases.¶

In the abstract, however, they each operate on the following pieces of information:¶

Subject:

The subject is the identity (individual or group/role) that intends to perform an action.¶

Action:

The action the subject intends to perform may be as simple as reading or writing a resource, but can be more complex.¶

Object:

Actions are performed on objects , such as a file or network resource.¶

Request Tuple:

A request tuple consists of the subject, action and (optional) object.¶

Privilege:

A privilege encodes whether or not an action is permitted.¶

Authorization Tuple:

An authorization tuple encodes system state, and is thus a tuple of a subject, a privilege and an (optional) object.¶

The act of authorization translates from a request tuple to a Boolean response determining whether a request is permitted. Similar to authentication above (Section 4.1), we can subdivide this into distinct steps that occur in phases.¶

Authorization Management:

The first phase is authorization management , in which an identifier is associated with authorization tuples according to the management scheme in use.¶

Authorization Query:

In the authorization query phase, a request tuple is used to look up in some store whether the request yields a positive or negative response. Conceptually, this resolves a request tuple into an authorization tuple -- or lack thereof. This effectively yields a boolean response to the request.¶

Access Granting:

The final stage is to grant access based on whether the authorization query step succeeded or failed.¶

Management of authorization always occurs in a preparation phase. Upon request for access to a resource, systems perform authorization query, immediately followed by access granting. The conceptual translation from request tuple to authorization tuple is an implementation detail, and typically remains conceptual at best.¶

By contrast, distributed authorization instead deals in authorization tuples, which can be stored and distributed out-of-band. Because of this, they can encode the least authority (see POLA, Section 3.5), independent of which mapping system was chosen to manage authorization policies.¶

It may be of interest that and authorization tuple is semantically equivalent to an RDF triple ([RDF]), in that it encodes a specific relationship between a subject and an object. Authorization tuples that consists solely of IRIs [RFC3987] are also syntactically RDF triples. This implies that authorization tuples can encode arbitrarily complex authorization information by building the knowledge graph resulting from resolving such an RDF triple.¶

5.1. Object-Capabilities (OCAP)

Dennis and Van Horn described an approach for securing computations in "multiprogrammed" systems in 1965/66 ([OCAP]). The context in which they operated had little to do with modern distributed systems.¶

However, they recognized the trend of running computing systems complex enough that multiple programmers would contribute to its overall function. This raised a desire for access control to individual sub-functions, which a security kernel within the operating system was to provide.¶

The key differentiator to other systems was that in OCAP, a calling process was to present a "capability" , a serialized token to the process being invoked. This capability was intended to encode all relevant information the called process would need to determine whether the caller was permitted to perform such an action.¶

These properties of being serializable and containing all relevant authorization information imply that, conceptually, capabilities are cached results of an authorization query . The called process can then perform access granting without issuing such a query itself, thereby temporally decoupling the two functions.¶

5.2. Identity-Capabilities (ICAP)

The OCAP system proved to have a particular weakness, namely that "the right to exercise access carries with it the right to grant access". This is the result the information encoded in an OCAP capability: it contains a reference to the object and action to perform, but does not tie this to any identity.¶

In 1988, Li Gong sought to address this with an Identity-Capability model ([ICAP]). Including an identity in the capability token arrives at the authorization tuple in Section 4.2.¶

Furthermore, ICAP introduces the notion of capability use in networked systems. ICAP does this by temporally decoupling the authorization query from access granting.¶

The main criticism levelled against the paper and capability-based approaches in general in the following years was that some functions were missing, such as a check for revocations. Proposals to address this often added centralized functions again, which led to criticism of the distributed approach in general.¶

5.3. Pretty Good Privacy

While we previously discussed PGP in terms of authentication in Section 4.1.1, a key property of PGP is the introduction of trust signatures ([RFC4880], Section 5.2.3.13).¶

Trust signatures do not merely authenticate a user, they introduce a kind of authorization as well, as they carry specific notions for what the provided public key may be trusted for. The trust signature thus encodes specific kinds of privileges of an authorization tuple , while the public key encodes a subject . The only component missing in the tuple is the object .¶

While the authorization tuple in PGP is incomplete, the system is based on public key cryptography, and can thus be used to securely verify a binding between the tuple elements.¶

5.4. JSON Web Tokens (JWT)

JSON Web Tokens ([RFC7519]) provide a standardized way for serializing access tokens. Current uses are in systems with centralized authorization functions such as OAuth ([RFC6749]).¶

However, the fundamental notion of capabilities, that a serializable token carries authorization information, is provided also here. Furthermore, JWT combines this with cryptographic signatures, providing for - in theory - temporal decoupling as previously discussed.¶

It's well worth pointing out that JWT is suitable as a portable, modern capability format - all it requires is to encode all necessary information within its fields. One serialization format in JWT for this is [UCAN].¶

5.5. ZCAP-LD

Aimed at the linked data space, [ZCAP-LD] is an expression of cryptographic capabilities for authorization that relies heavily on linked data. While conceptually, the specification shares many similarities with the capability concept in this document, the use of linked data can lead to systems that do not provide for temporal decoupling .¶

Linked data has the downside here that data relationships may need to be resolved at the time of access granting , thus effectively re-introducing parts of an authorization query again at this point.¶

The concept of distributed systems underlying linked data thus differs fundamentally from the one in [RM3420]. Where the former treats distribution as distribution of concerns across different services providing parts of the linked data set, the latter is more concerned with resilience, specifically how to continue operating in the (temporary) absence of such services.¶

5.6. SPKI Certificate Theory

Cryptographic capabilities are not a new concept. [RFC2693] provides a thorough analysis of how they may be modelled in [X.509] certificates, and includes a discussion on various methods for modelling authorization that also draws comparison to PGP above.¶

The two major drawbacks of this document are:¶

1. Being a thorough theoretical discussion, it is dense. It does not provide an easy path for implementors to design capability based authorization systems; as a reference, however, it is indispensable.¶

2. With its focus on X.509 certificates as the encoding and transport mechanism for authorization material, it does not lend itself well to a number of use cases; this may be why overlapping proposals exist.¶

   Recall how DTLS [RFC9147] contains provisions for handling large certificates, and that JWT and ZCAP-LD provide serializations more suitable to web technology. Arguably, a discussion of the principles of authorization removed from the specifics of X.509 are required.¶

In the light of that RFC, this document focuses on relative simplicity and freedom from specific serialization formats for capabilities.¶

5.7. Macaroons

Capabilities bear resemblance to Macaroons, presented in 2014 [MACAROONS]. The Macaroons paper outlines a system of caveats under which access may be granted to the bearer of a macaroon token. Macaroons are secured using an HMAC. The paper also lays out the potential use of public-key based security for Macaroons.¶

Macaroons may be viewed as a specific implementation of the concept of capabilities with the focus of providing authorization in cloud environments. The paper positions the solution primarily as adding finer granularity to bearer tokens. In doing so it lays out a more complex caveat scheme than this document, which excludes the specifics of the privileges that may be issued.¶

5.8. CapBAC

The CapBAC system [CAPBAC] is focused on IoT applications, and in particular attempts to solve the complexity problems in this application domain. The paper is predominantly concerned with mapping abstract authorization concepts into specific privileges encoded in a capability. It is thus an excellent companion and motivation for this document.¶

The focus of this document, by contrast, is on mapping out the minimal necessary components of making capabilities workable.¶

5.9. Power of Attorney

The oldest kind of prior work in this field is the concept of Power of Attorney, as exercised throughout much of human history.¶

In a Power of Attorney system, an authority (a king, etc.) grants a token (an official seal, ...) to a subordinate which makes this subordinate recognizable as carrying some of the king's powers and privileges.¶

Modern day Power of Attorney systems abound, and may be formalized as notarized letters granting such and such rights to other people.¶

Capability-based authorization schemes are no different to this kind of system in principle. In both kinds of systems, the token itself encodes the privileges of the bearer. [POA-IOT] describes such a system for the Internet-of-Things.¶

6.1. Grantor

A grantor, sometimes called principal, has authority over an object, and generates authorization tuples for use in the overall system.¶

As we describe cryptographic systems, a grantor is represented by an asymmetric key pair. Endowment for a grantor is out of scope of this document; for the purposes of distributed authorization, the grantor key pair is the grantor.¶

6.2. Agent

The agent is the element in a distributed system that executes a requested action after verifying a capability. It typically manages objects itself, or provides access to them.¶

6.3. Verifier

The verifier is a role in the system that verifies a capability. While verifiers can exist in a variety of system nodes, it's a mandatory part of the agent role.¶

Outside of the agent, verifiers may exist in intermediary nodes that mediate access to agents. An example here might be an authorization proxy that sits between the public internet and a closed system. While it may not be an agent in and of itself, it can still decide to reject invalid requests, and only forward those to agents that pass verification and its own forwarding rules.¶

6.4. Time-Delayed Transmission Channel

We introduce the concept of a time-delayed transmission channel to illustrate that communications between grantor and verifier is not possible in real-time.¶

This is not fundamentally different to the delay between the preparation and real-time phase employed so far in this document, except to add the requirement that the capability is transmitted at some point during this delay.¶

In practice, this could be at the beginning of the delay, or when the access request is made. They key point, however, is that this transmission of the capability encompasses all the communication that occurs between the functions in the real-time phase, and those in the preparation phase, and thus decouples these functions in time.¶

6.5. Grantee

The grantee is the entity to which a privilege is granted.¶

A grantee SHOULD also be represented by an asymmetric key pair in order to perform distributed authentication.¶

6.6. Object

An object is a resource the grantee wishes to access. This can be a file, or a networked service, etc.¶

6.7. Privilege

A privilege encodes whether an action (on an object) is permitted (for a subject); see {#sec:authorization} for an explanation.¶

For the purposes of creating capabilities, a privilege must have a canonical encoding. The semantics of each privilege are out of scope of this document, and to be defined by the systems using distributed authorization.¶

That being said, a typical set of privileges might include read and write privileges for file-like resources.¶

6.8. Validity Metadata

In practical applications of distributed authorization scheme, validity of a capability may be further scoped. We already discussed the need to scope it to an authorization tuple, but further restrictions are likely desirable.¶

For example, a set of not-before and not-after timestamps exist in e.g. [X.509] certificates; similar temporal validity restrictions are likely required in practical systems.¶

However necessary they may be in practice, however, such additional validity metadata has no bearing on the fundamental concepts outlined in this document, and is therefore considered out of scope here.¶

6.9. Capability

A capability provides a serialized encoding of previously listed elements:¶

1. Fundamentally, a capability MUST encode an authorization tuple, consisting of:¶

   1. A subject identifier.¶
   2. A privilege. ¶
   3. An object identifier.¶
2. A grantor identifier MAY be required in order to identify the grantor key pair used in signing and verification.¶
3. Validity Metadata SHOULD be included in practical systems.¶
4. In order for a verifier to ensure the validity of a capability, it MUST finally contain a grantor signature over all preceding fields.¶

The authorization tuple permits an agent to determine what kind of access to grant or deny. The grantor identifier provides information to the verifier about key pairs used in the authorization. While the signature proves to the verifier that the grantor did indeed authorize access, the validity metadata limits access to whichever additional scope the grantor decided upon.¶

6.10. Authorization Process

Having identified the elements, we can now describe an abstract process in a distributed authorization system.¶

The process is split into the two preparation and real-time phases.¶

In the first phase, the grantor issues an authorization query (((authorization query))) to an authorization tuple store, which stands in here for the specific process by which authorization is managed, and produces tuples. Based on the response, it serializes a capability and adds its signature over it.¶

The capability then gets transmitted via the time-delayed transmission channel to the second phase, providing temporal decoupling between the phases.¶

In the second phase, the grantee requests access to some object from the agent. The agent must send a verification request to the verifier (which may be a subroutine of the agent; no network transmission is implied here). The verifier responds by either permitting access or not. If access is permitted, the agent grants access to the grantee. Because the capability encodes all required information for the verifier to perform this step, it does not need access to the authorization tuple store itself.¶

Note that the capability can be transmitted to any entity in the second phase; all that is relevant is that it ends up at the verifier. If it is transmitted to the grantee, it has to pass it on to the agent as part of the access request. If the agent receives it, it has to pass it on to the verifier as part of the verification request.¶

8.1. Human Rights Considerations

This document lists human rights considerations as a use case, see Section 3.6.¶

8.2. Protocol Considerations

There are no specific protocol considerations for this document.¶

However, protocols transmitting capabilities MAY provide some relief to human rights concerns Section 3.6, e.g. by providing confidentiality via encrypted transmission.¶

8.3. Security Considerations

This document does not specify a network protocol. In fact, it deliberately requires no specific protocol for transmitting capabilities. As such, much of [BCP72] does not apply.¶

However, distributed authorization does not require the invention of new cryptographic constructs; the document is deliberately phrased such that the choice of such constructs remains implementation defined.¶

As such, some security considerations are supported by the use of capabilities for distributed authorization, such as preventing unauthorized usage and inappropriate use.¶

Some notes on specific considerations follow.¶

8.4. Privacy Considerations

As part of supporting human rights considerations as a first class use case, exploring privacy considerations as covered by [RFC6973] is worthwhile.¶

In particular, distributed authorization schemes address the concerns of: intrusion and misattribution (as related to pseudonyms only).¶

It's also worth highlighting that surveillance and stored data concerns, as well as disclosure, are not addressed. In order for distributed capabilities to work, any likely recipient needs to be able to decode them.¶

The threat model then assumes that all capability data is accessible to anyone, which is why the use of pseudonymous public-key based identifiers is suggested. Sufficient care must be taken in key rotation, etc. in order to provide additional protections.¶

Note that despite this, nothing prevents a system from encrypting capabilities for use only by a single authorized party, which means that these last concerns can be addressed in the surrounding system.¶

8.5. IANA Considerations

This document has no IANA actions.¶