Work in progress This is work building up to a final requirements document on grid AAA/Security. It draws very greatly upon [wiki:RequirementsBibliography Shawn Mullen et al's] (2004) document on "Grid Authentication, Authorization and Accounting Requirements". Most of the requirement text is from this document and where we disagree, we have indicated this

Thus:

REAL TITLE: The requirements for an ideal grid

Should this be "an 'ideal' grid" or "an idealised grid" or whatever?

This document

The sections of this document are:

  1. Abstract
  2. Terminology and definitions
  3. Grid use models
  4. Site Authentication Requirements
    • (Terminology and definitions)
  5. Site Authorization Requirements

    • Unfinished list - make it better later! (Could also use some internal anchors.

Abstract

To be written last

However, it should include some definitions of "must" and "should" etc. AND a discussion of the grid model (e.g. Customer-Service, Technical Research User (agnostic of node), Technical Research User (node specific), Privacy/Conf./High security etc. (The musts and shoulds will be different for each grid model)

The definitions of must, should, should not etc. are as defined in RFC2119 (Bradner 1997).

Terminology and definitions

To be written late-on when we know what we need.

Grid use models

Introduction

The use models of grids are written about elsewhere by this project. However, for clarity, they are reproduced here in summary form.

BLAH BLAH

(Chapter 1) Site Authentication Requirements

Terminology and definitions

"User secrets" refers to values intended to be known only by the user, known by the user and an authentication infrastructure, or known only to an authentication infrastructure and employed on the user's behalf after the user has authenticated with some other secret(s).

To sidestep such questions as whether "a day" means eight hours or 24 hours and just how long a month is, we will deal in seconds but not quibble over implementation variances at the 10% or 20% level.

Credentials are assumed to have lifetimes which bound their period of validity. "Long-lived" credentials have lifetimes of 1,000,000 seconds (1 megasecond or 1 Ms) or more. "Short-lived" credentials have lifetimes of 100,000 seconds (0.1 Ms) or less. Lifetimes between those limits are "intermediate." The terms long-lived and short-lived may also be applied to the secrets employed by a user to acquire credentials, although the only short-lived user secrets known to be commonly employed are one-time (or "single-use") authenticators.

(Conversions: 0.1 Ms is a bit more than a day; 1 Ms is a bit less than 2 weeks.)

If a credential's lifetime can be extended by the user, using no more proof of identity than the credential itself, this is considered "renewal" of the credential, while if the process of extending the lifetime requires measures equivalent to those employed in its initial acquisition, we consider the result a new credential.

We specifically do not consider "post-dated" credentials -- those with lifetimes that begin at some point later than the time of the authentication act. Neither do we consider the relative strengths of cryptographic protocols, algorithms, and key lengths. We assume they are always designed, selected and implemented appropriately.

Identity

Sites will often make authorization decisions on an aggregate basis: on Virtual Organization (VO) membership or group membership. However, at times it may be necessary to set access rights at the granularity of a single user. Sites therefore may need to reserve the right, and preserve the ability, to set authorization at this level. Incident handling requires the ability to identify the legitimate owner of credentials presented during transactions under investigation. However, this may be done in concert with a trusted partner (i.e. the site could accept a pseudonym with good provenance but later need to know who/what is the actual owner of the credentials presented, and could co-operate with that partner to determine the identity of the end entity in question). 

IMPORTANT REMOVAL FROM THE MULLEN et al DOC:
"Accordingly, every set of authentication credentials should be tied to the identity of an individual..."

Also minor changes in the next paragraph (shown in italics).

Accordingly, every set of authentication credentials should be traceable to the identity of an individual, because this provides stronger security by way of auditability, revocation, and problem determination. However, in very special cases there may be occasion to forfeit these benefits in order to provide temporary and generic identities.

For example, an Internet cafe could provide temporary (very limited lifetime) credentials authorizing use of grid resources based solely on the fact that access was purchased. Such an identity may be a psuedonym such as "Customer 24." However, please note that it may be better to achieve a solution where a customer-service provider (CSP) model [see above xxxx] is used in such a situation.

Where credentials are supplied pseudonymously by an identity provider (home organisation) to a grid node or nodes, a service level agreement must be in place between that identity provider and the grid community that will ensure the rapid revocation of those credentials when demanded. This may be invoked by a service provider (or grid node) that detects a security problem associated with those credentials. It may be advantageous that credentials may thus be issued and re-issued easily, and therefore the consequences to the user (or end entity) may be minimised following a false positive (i.e. mistaken) revocation.

May need to re-write the above as it is rather wordy and verbose and, erm, wordy.

Other, similar identity indirections are expected:

Secure anonymous communications may still be allowable, and appropriate, for functions that do not require user authentication.

For example, in the case above of cafe access to Grid resources; the user may still require secure conversation because the results of the data derived may have some proprietary value. Anchor(assurancelevel)

Assurance

An authentication system may provide multiple methods for a user to perform their initial authentication, and these methods may differ in their convenience, resistance to attack, and risks of exposure of secrets. Even when an implementation offers its users only one method, it may not be clear to relying parties which method it is. Since some inverse correlation does exist between convenience and strength of authentication, there may be inducements to allow and employ multiple levels of authentication if sites make some class of services available through weaker but less burdensome authentication methods. 

  V Many changes (unless highlighted) to the Mullen et al doc exist in the following sections V

In a deviation from the Mullen et al document, we suggest that there is certainly a greater variety of assurance levels possible, due to both the nature of the authentication tokens (and their storage/encryption) and the policies and procedures regarding their issue and timely revocation. Thus, a numbered scale of assurance level should exist and a value should be passed from the identity provider (home organisation) to the grid node or nodes for short term credentials, or should be kept permanently (implicitly or explicitly) within longer term credentials. This assurance 'grading' needs to be of a value and format recognised by an international standard or (more likely in the shorter term) be based upon a system agreed upon by collaborators within a particular grid.

The system for grading the assurance level of each authentication assertion is beyond the scope of this document. However, this assertion should be made and transmitted. If practicable, the method used to perform authentication should be deducible from credentials, but this requirement is secondary to the requirement of the transmission of assurance level.

Levels of authentication strength

We consider this to be eyond the scope of this document. However, we demur from Mullen et al. in that we believe that the concept of "authentication strength" is too close to the concept of "assurance level". The reliability or trustablility of an authentication event or authentication token is based equally upon the technology used (and short/long term credentials and encryption etc.) and the policies used to initially authenticate, maintain data, and renew and revoke credentials.

Mode of storage

We recognize the following modes of storage of users' long-term secrets (whether used directly to authenticate to a grid resource or to a proxy), each with its own set of vulnerabilities:

What you know

What you have

Ranking of storage modes

We agree broadly with the Mullen et al document, but believe that this is where home organisation (or "identity provider") security policies play a role in ranking the storage mode in a way that would affect a judgement of assurance level.

It is not possible to give a strict ranking of storage modes discussed in section 1.5.3.3 relative to safety without asking and answering a number of questions about the details of the secrets, their storage, and their registration as the users' authentication information. Also, users may perform unsafe actions (knowingly or unknowingly) which place their secrets at much greater risk of disclosure.

Deducible Authn strength

  ^ END of Many changes (unless highlighted) from the Mullen et al doc ^

There are a number of cases where processes running on a machine need to authenticate to other processes. Automated processes may have to act as authenticated clients and users may wish to have automatic software ("cron jobs") that require automatic authentication. All of these should be somehow restricted such that theft of credentials from an individual machine does not easily permit their reuse elsewhere. In either case, secrets will be of the "stored" class and must be considered to be stored in cleartext form, regardless of any measures which obfuscate them.

Lifetimes

All forms of digital credential in common use are subject to possible theft and misuse. The probability of such an event is monotonically nondecreasing with time. The countermeasures against eventual credential theft are expiration and revocation. Neither measure alone is sufficient to prevent all misuse, nor is the combination of the two.

Two types of digital credential should be highlighted here to take into account the roles and behaviour of proxy parties. In many use-cases it has been found to be necessary to generate proxy credentials. These may differ in digital nature and in assurance level from the original user digital credential. They are typically shorter lived than the original digital credentials, but there may be exceptions to this generalisation. Thus, we hereby draw out the two concepts of user credential and proxy credential. 

V Much changed in the following sections from the Mullen et al document to accommodate user and proxy credentials and to make some of the timings more onerous/rigorous V 

Note: In the next paragraph, I changed "authority parties" (which was undefined, I believe) to "trusted members of the grid community".  Also, beefed this up a bit, as the Mullen et al doc. said that 3rd-party requests for revocation were not time bound.  They really should be, and therefore I removed quite a bit of original text from "...under some circumstances" to the end of the paragraph .  N.B. The Mullen et al. doc seemed to suggest that in some circumstances RAs should able to demand an immediate revocation: in the DCOCE findings, we thought that these people should be able to issue revocations more freely than most other parties.

The lifetime of authentication secrets is a separate parameter from the lifetime of credentials. 

N.B. The above is original text.  However, is this skewed in support of X.509 certs?  i.e. username/password combinations need to be considered as single objects.  (I'm not sure, but) I suggest changing the text to:
The lifetime of authentication secrets is a separate parameter from the lifetime of the digital credentials (even where the combination of secret and public credential is used in some way to form an authentication credential). 

All of the above does not talk about proxy authentication credentials (except for what we have added in the 'lifetime' section).  In the Mullen et al doc., these are considered in the authorization section (i.e. later).  However, we will need to include them here somehow (as well?, instead?) as they are, arguably, relevant to X.509 and, definitely, Shibboleth.  Nevertheless, it may be worth debating whether proxy authN credentials exist at all (i.e. once an entity is authenticated, everything that follows is authZ???). 


(Chapter 2)  Site Authorization Requirements




Terminology


Terminology used in this document strives to be consistent with that used in the Authorization Frameworks working group REF? xxxx. 




"User" is a synonym for end entity and for subject used in the more general framework document. We preserve the use of "user" since it is more widely used within the site operations community. 




"Groups" refer to groups of end entities which are accorded equivalent rights for purposes of obtaining a particular set of privileges. 




"Role" refers to the set of attributes an end entity is presenting with a particular request for obtaining or asserting a privilege. 




"Provenance" refers to information about the history of a request or of any type of assertion. Examples include: the identity of the original requester; and the identity of the entity that is making the assertion. 


Authorization Process


V We take an important and different view from Mullen et al, expressed between these marks V 




A VO must manage information regarding users that can be used for authorization decisions.  This information may be made available to certain trusted third parties.  Thus, a typical authorization process may have several steps (for example, but not necessarily in the following order, user authorization, VO authorization, site authorization, resource authorization) with various implementations. 
Users and VO managers must be able to rely on consistent interpretation of their policies. 




The Virtual Organization must be able to decide user membership policy and allow sites to set user authorization policy.  However, it is likely that a degree of co-operation between the VO and the site will be desirable in setting the site's authorization policy in most cases. 
^ We take an important and different view from Mullen et al, expressed between these marks ^ 




The authorization method must be application independent. 




Mutual authorization may be required. 
An application or end entity may need assurances that the resource is authorized to run a specific job.  The distributed program or grid job in and of itself may be of value.  The results may be of value and need protection from dubious resources. 
A grid job may need to specify that it is only run on systems with security level B operating systems, or systems not directly connected to the Internet, or some other operations requirement. This is more relevant in the OGSA model where service factories may incorporate more resources to handle service request loads. 




Maintain Provenance 
The authorization mechanism must preserve the Subject Identity of the user who originated the request, except in transactions taking part within the Customer-Service model of operations: in this case, the grid service is trusted to have carried out the authorization check on the customer and to be acting on his/her/its behalf. 




Provide for method of grouping users 
It should be possible to assign a user to a group. The authorization of resource access can be managed by managing permissions of the group. 




Authorization Level Dependent on Authentication Strength 
The authorization for access to a resource at a particular level may depend on the strength of the authentication. The level of authentication or, more likely, the level of assurance (agreed upon by the grid community, see [wiki:RequirementsDoc Section 5.3]), must be included with the credential information presented to all resource managers. 




Call-outs 
Call-outs prior to access to resources may be provided as a form of authorization control by the virtual organization, the site(s) and each resource provider. 




Revocation 
There must be the ability to quickly revoke a particular remote authorized service that may be operated under dubious procedures. The timescale for this revocation should be of order 0.1 Ms. 
For example, if a remote processing resource steals computation results, it should be removed from the directory of processing resources.  This is difficult in the context of the current Grid technology because of the open resource registration process and aggressive discovery algorithms.  Similar such directory services on the Internet have a history of exploitation. 


Authorization Attributes






Attribute Authorities 
In expected grid operations, authorization attributes are managed by authorization authority servers run by VOs, by sites or other authoritative entities. These authorization attributes may contain specific authorization privileges, indicating to sites that they should be authorized to act in a particular role, or may contain statements of membership in a particular group within the VO. 
Unfinished....