ABAC access control for Fedd

The ​ABAC access control system, developed at Stanford and realized by a group in Trusted Information Systems that later moved to SPARTA, is a formal, flexible, scalable access control system based on formal derivation of user attributes, attested by other trusted users. We have been ​planning to integrate it into fedd for some time. We have recently implemented ABAC in a ​portable library and this implementation has been integrated into fedd. Under our associated TIED project, we have completed a similar ​integration with GENI's ​reference aggregate manager, part of their developing ​GENI API.

This page describes the use of ABAC with fedd, concentrating on using the transition tools to create initial ABAC credential stores from which to run fedd. In order to get the most from this page, you should be familiar with

• three-names and fedids
• ​The ABAC model as it is used by DETER.
  • There is a ​worked example that can be helpful here
• The existing fedd access databases

Storing Credentials

Users and servers (fedd instances) now both have credential stores to maintain. When using ABAC credentials to enforce the same kinds of access control as three-names neither users nor servers will see much change. Credentials will be managed transparently.

By default a user will maintain a credential store in a directory named .abac. Any credentials in the users abac directory will be passed as parameters when requests are made to fedd. The user can limit the credentials passed by specifying a directory with only a subset of the credentials.

Similarly, a running fedd instantiation will be configured to use a specific abac directory in which its credentials are kept. As the instance learns more credentials they are aslo stored in this directory.

Unlike a user, the fedd directory contains metadata beyond the credential lists. This directory is populated using the access_to__abac.py command for access controllers and the fedd_to_abac.py command for experiment controllers.

Operation of Experiment Controller

The experiment controller can operate in two modes, as an experiment controller dedicated to use by a single user, or as a service on behalf of multiple users. In the first case the controller acts as the same principal as the user commands. Essentially the experiment controller is a command run by the single user. In the second case a common piece of infrastructure acts as an experiment controller for many users. Because this fedd acts as a distinct principal, some additional delegation is required to make the access control reasonable.

When the experiment controller is run as a service, it acts as a separate principal for the purposes of experiment creation, and the user and the controller negotiate what principal the controller will act as when making resource allocations from access controllers. This allows a user to delegate a subset of attributes to the experiment principal and then allow the experiment controller to act as that principal. (The experiment principal may have attributes delegated by multiple principals, if needed). This extra principal is required to prevent the experiment controller from combining rights from multiple principals that did not intend it. The figure below shows this:

When a user is running an experiment controller on their own behalf, like a utility program, the situation is much simpler. Both experiment controller operations and experiment allocations are simply carried out as the user. The user's fedid is specified for both the experiment controller's ID and the experiment ID, leading to the simpler situation below:

Expressing Three-names as ABAC credentials

Three-names are really an encoding of a fairly small attribute space used by access controllers to control which fedids can perform operations and how to map authorized fedids into the local identifier space. Experiment controllers assert the attributes.

The attributes an experiment controller asserts are given in its access database in lines of the form:

fedid:xxx -> (proj1, user1)

That line asserts that this controller will assert that a user with fedid of xxx is in project proj1 and is user user1. This encodes in ABAC (using the RT0 syntax supported by ​libabac as:

ec_fedid.user_user1 <- fedid:xxx
ec_fedid.project_proj1 <- fedid:xxx

ec_fedid is the fedid of the experiment controller. The fedd_to_abac.py script converts an experiment controller's access DB to an ABAC credential store with those credentials, encoded as X.509 attribute certificates.

The access controller uses three-names to both make the access control decision (does a requester have the right to make this call) and a mapping of the fedid into the user access space. For example a line like:

(ec_fedid, proj1, user1) -> create, (DETER, faber, faber)

says that a user with a project of proj1 and a user of {{user1}} attested by the controller with fedid ec_fedid can create experiments as local user faber on the local DETER project. The ABAC credentials for the access control decision look like:

ac_fedid.project_proj1_user_user1 <- ec_fedid.project_proj1 & ec_fedid.user_user1
ac_fedid.create <- ac_fedid.project_proj1_user_user1

The first rule names the conjunction of two experiment controller attributes and the second line maps that to a local create attribute. The first attibute is used to map to local credentials.

The mapping to local credentials is accomplished by a map file that maps an ABAC credential to a local data tuple. The encoding above would be:

ac_fedid.project_proj1_user_user1 -> (DETER, faber, faber)

The tuple has the same format as the non-ABAC access control database for that kind of controller.

ABAC Logic: RT0 vs. RT1

ABAC can support several underlying logics. The current logic, RT0, implemented by ​libabac represents attributes as simple strings. This is why encoding a project takes the form project_name.

RT1 allows for parameterized attributes, that is an attribute is a string and 0 or more typed parameters. Those attributes would take the form of project(name), and would allow some more sophisticated deduction rules. While the RT0 vs. RT1 distinction is only a convenience for the encoding of three-names, we intend to pursue implementing the more powerful logic.

Conversion Utilities

This section describes the utilities for converting existing three-name-basedconfigurations into ABAC. Users do not need to do anything, but there is a utility for converting experiment controllers and access controllers.

In addition to running the utilties, add set auth_type to abac and auth_dir to the ABAC directory created by the utilities below. For access controllers, the accessdb variable should point to the generated map file.

fedd_to_abac.py

fedd_to_abac.py converts an existing fedd experiment controller configuration into an ABAC configuration. A certificate and key are required as well as an output directory for the ABAC authorizer (target of the auth_dir parameter in the configuration file.

It takes the following arguments:

--cert=file

A file containing an X.509 certificate in pem format. This is the identity that will assert the attributes and should probably be the same as the experiment controller's identity. It may be a combination certificate and key file.

--key=file

The key for signing attributes. It should be a pem file with an RSA key in it. This can be omitted if --cert specifies a combination file.

--dir=dir

The output directory for the authorizer.

--make_dir

If given, make the output directory.

--debug

Just output the creddy commands that would be issued to create the certificates.

--policy_only

Only output the ABAC certificates, not the full ABAC authorizer. This is primarily for debugging.

--update

Update the output directory rather than creating it. This allows the authorizer to remember credentials it has learned or issued while updating the policy.

After the named arguments are given, the aceess DB file is a required argument.

access_to_abac.py

access_to_abac.py converts access controller policies into ABAC and generates a DB for mapping the derived attributes into an appropriate credential mapping database, the target of the accessdb configuration directive. Keys and certificates are required parameters as well as a directory for the ABAC authorizer and a file for the access DB.

It takes the following arguments:

--cert=file

A file containing an X.509 certificate in pem format. This is the identity that will assert the attributes and should probably be the same as the experiment controller's identity. It may be a combination certificate and key file.

--key=file

The key for signing attributes. It should be a pem file with an RSA key in it. This can be omitted if --cert specifies a combination file.

--dir=dir

The output directory for the authorizer.

--type=type

The type of access controller, emulab, dragon, etc. Used to parse the various formats of the old accessDB.

--quiet

Do not produce extra output

--no_create_creds

Do not create the authorizer (debugging only)

--file=old_accessdb

The accessDB to convert. May also be specified as a bare parameter after the named parameters have been given. {{{--mapfile=}}new_accessdb:: The new accessDB to create. It must be a different file name than the file being converted.

--no_delegate

By default the ABAC created attributes include a layer of delegation for a standalone fedd running with a different principal. Specifying this parameter eliminates a layer of delegation. Even for single-user experiment controllers, this option is rarely necessary.

--no_auth

Do not create an authorizer. For debugging only.

--debug

Just produce debugging output.

If --file is not given, the accessDB to convert must follow.