OASIS Mailing List ArchivesView the OASIS mailing list archive below
or browse/search using MarkMail.


Help: OASIS Mailing Lists Help | MarkMail Help

xacml message

[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]

Subject: Re: [xacml] Attributes of Relations

Hi Erik,

On 5/01/2013 12:05 AM, Erik Rissanen wrote:
Hi Steven,

Sorry, but I am still not convinced. ;-)

You say that nothing more needs to be specified. I still doubt that is true, and we'll end up making it more
and more complex over time as customers come with use cases we did not anticipate. Just to throw one in
here, what about if in the example which is quoted below, there are multiple relationship objects,
representing the employment history of the subject, each one dated, and you want to check that the
relationship which has the latest date is the one in which the subject is employed by the ip owner? Can you
do this with the iterator?

Yes. Add the following term into the arguments for the "and" function.

    <!-- Return true if this subject-to-organization-relationship,
         i.e., $a, has the latest start-date. -->
    <ForAll VariableId="$b">
      <!-- $b is bound to each relationship-ref URI in turn -->
      <Apply FunctionId="any-of-all>
        <Function FunctionId="date-greater-than-or-equal"/>
        <Apply FunctionId="attribute-designator">
          <VariableReference VariableId="$a"/>
          <AttributeValue DataType="anyURI">start-date</AttributeValue>
          <AttributeValue DataType="anyURI">date</AttributeValue>
          <AttributeValue DataType="boolean">false</AttributeValue>
        <Apply FunctionId="attribute-designator">
          <VariableReference VariableId="$b"/>
          <AttributeValue DataType="anyURI">start-date</AttributeValue>
          <AttributeValue DataType="anyURI">date</AttributeValue>
          <AttributeValue DataType="boolean">false</AttributeValue>

Note that I continue to assume that all attributes can be multi-valued. If some
can assumed to be single-valued, then some of the iterators can be removed.

I am also worried about the learning curve for the proposed new expressions.

It's minor compared to having to additionally learn SQL to configure attributes in
a PIP when writing XACML policies in the PAP, which is what you are proposing.

> If you look at the example,
there are many levels of indirection and I think that many users would struggle working with these kind of
expressions. Especially comparing this with a pretty simple plain SQL select to do what you did with the

A PAP implementation is free to present XACML expressions to users any way
it likes, including in a compact, SQL-like syntax that gets translated into
the highly verbose XML format behind the scenes. Just because the TC has chosen
XML as the syntax for representing XACML expressions doesn't mean that users have
to see it that way.

Anyway, I think the best way for you to proceed is to ask TC admin for a template for a specification
document and start editing the profile you are proposing, and it will be clearer if there is any interest
for it among users and vendors.

I intend to do that, but I'm not likely to get around to it before the RSA Interop.
Maybe you could do the same for what you are proposing.


Best regards,

On 2012-12-24 00:07, Steven Legg wrote:

Hi Erik,

On 20/12/2012 8:39 PM, Erik Rissanen wrote:
Hi Steven,

Regarding a PIP to query the context handler, it is something which is not really made explicit in the XACML
architecture diagram due to lack of granularity, but it is implied. One issue is that what one means with
the term PIP. Is that the remote data source, that is, the actual SQL database, only? That one of course
does not query the context handler, but the "PIP connector", which is part of the context handler perhaps
depending how one would elaborate the finer details in the XACML architecture, certainly has access to the
attribute values of the abstract request context. It will use these as key values in its queries to the
database, or it would be entirely incapable of the most basic remote attribute value retrieval.

Anyway, if you want something standardized, I would instead of re-inventing the wheel, look into defining a
query interface from XACML to something like RDF, which is capable of expressing these types of relations
and has a powerful query language and implementations are available, both commercially and open source.

I have no interest in standardizing anything like this because it splits
the policy logic over multiple systems and multiple forms of expression.
Unless you can find someone willing to do the work, you are left proposing
a solution for IPC that is not interoperable.

I might be mistaken, but my gut feeling is that you underestimate the amount of stuff needed if you consider
only this one use case. Once you have multiple different relations which are interconnected and you want to
match on several fields in different ways and do joins between the relations, an iterator just is not

What I have proposed already has the expressive power of joins over multiple
relations. A join is a selection over the cartesian product of a group of
relations. XACML expressions can do the selection. Even the any-of-any
function can create the product, though it only allows the selection to be
one of the predefined functions. The iterator expressions can do the product
*and* allow the selection to be any arbitrary expression. An attribute
designator as a function allows the product to be defined over any number
of linked objects, i.e., "relations". The example I gave had the equivalent
of five relations. So yes, you are mistaken. Nothing more needs to be specified.


Best regards,

On 2012-12-20 04:36, Steven Legg wrote:

Hi Erik,

On 19/12/2012 7:03 PM, Erik Rissanen wrote:
Hi Steven,

Yes, that is my proposal. See inline for responses to your concerns.

On 2012-12-19 00:44, Steven Legg wrote:

Hi Erik,

On 18/12/2012 2:18 AM, Erik Rissanen wrote:
Hi Steven,

My point in my email was that you cannot just flatten attributes like that, as you describe yourself,
also that I don't think it is worth doing in XACML.

One option is to extend XACML. There are multiple ways to do that. One is to define a tuple/relation
datatype and operators on it, which I think is a relatively nice and clean solution. Another one is
what you
propose is to iterate over categories, which I don't quite like as much since it appears more kludgy
well established relational algebra on tuples. We could debate this...

However, you can also handle a requirement like this in a PIP, in which case you do not need any
changes to

Given that, we could define these extensions into XACML, but I have doubts about what the value of this
would be since it complicates the XACML language, making XACML harder to learn and implement. And since
these operators/iterators are already available in commonly used PIP backends such as SQL and LDAP,
so the
XACML industry would essentially be re-inventing these features into the XACML language, although every
organization already has easy access to solve these use cases using the existing XACML spec and PIPs.

You really need to describe your thinking in more detail.

Since the PIP only supplies attributes, my best guess is that you expect to
use XACML attribute values, typically of the boolean data-type, as surrogates
for the result of SQL statements or LDAP queries evaluated internally by the PIP.
For example, an attribute called subject-employed-by-ip-owner that indicates the
boolean result of an internal PIP query to evaluate whether the subject is an
employee of the IP owner. I see a number of problems with an approach like this.

(1) It assumes that the PIP has all the necessary information to query. Some
attributes (like ip-owner) may only come from the PEP in the authorization request.
It is only the context handler and PDP that are guaranteed to have the complete
request context for evaluation.

Typically in an implementation a PIP can ask the context handler for more "key values" which it would
so this is not a problem in practice given that the implementation is done right.

Is there a standard somewhere that describes how a conformant PIP can query the
context handler ? I haven't noticed one, either formal or defacto. More importantly,
are there any interoperable implementations that can do this ? I'm not aware of
any. The essence of your solution seems to be to ask implementors to do something
proprietary, which isn't a very satisfying answer for a group that is supposed to
be about enabling interoperable implementations.

(2) It is necessary to define a new attribute and configure the PIP with a new
SQL statement, LDAP query, or whatever, for each new expression the policy writer
wants to use that is beyond the current capabilities for XACML to express.

That is correct. I don't see this as a huge problem.

There is no XACML standards support for doing this sort of PIP configuration
and coordination between PAP and PIP. I see that as a huge problem for

(3) It splits the policy evaluation between the PDP and the PIP. It is no longer
sufficient to examine the XACML policies to understand the effect of access control
policy. It is necessary to also examine the SQL statements, LDAP queries, etc. to
understand what the complete authorization policy is. For interworking, it is also
necessary to export the XACML policies from the PAP and the SQL statements, LDAP
queries, etc. from the PIP.

Yes, this is the main drawback in my opinion. But it's something one can live with.

I'm not willing to live with it. Again there is no XACML standards support for
doing all this stuff. If you were offering to create those standards it would be
a different story, but it would be a lot more work than the extensions I was
talking about. And I would be the one writing up those extensions anyway.

If that isn't the approach you had in mind, then you need to describe what you
are thinking. An example of how you would tackle a condition like "the subject is
an employee of the IP owner" would be useful for comparison.

This is the approach I do have in mind.

The reason I prefer this is because the amount of "new stuff" we would have to add to XACML to handle all
such cases of attributes of relations is a big deal.

You overstate the amount of new stuff. The non-trivial example I gave had only
three lines that weren't core XACML, and one of those was an end-tag. The iterator
expressions are a generalization of the numerous bag functions that already
exist. The attribute-designator is just an AttributeDesignator expressed as
a function. They are modest changes that provide a big improvement in
expressibility and solve the problems of the IPC profile, among other things.

> We will basically end up re-implementing a SQL like
language within XACML, and I doubt what the value of that would be since SQL implementation are ubiquitous
and can already do all this for us.

Adding something SQL like within XACML is going to be costly in terms of TC effort, implementation effort
and customer learning curve for XACML products.

There's already a big overlap between what SQL and XACML can express, they
just do it differently. XACML is already way more expressive than an LDAP
filter. It isn't necessary to reinvent SQL or LDAP filters to solve the
problems of the IPC profile. Much of the reinventing has happened already.
Adding a bit more to what's already there is enough.

> And in the end, it won't do anything which cannot be done
today, with the exception of a more fully visible authorization logic in the XACML policy.

That in itself is a good thing since it's one of the claimed benefits of centralized
authorization, but it also enables interoperable implementations. If that's not
what the TC is about, then what is it for ?


Best regards,


Best regards,

On 2012-12-17 07:43, Steven Legg wrote:

Hi Richard,

On 30/11/2012 8:13 AM, Hill, Richard C wrote:
The topic of "attributes of relations" came up during our work on the IPC profile that I would like to
to the list for discussion.

For example, an intellectual property agreement (Copyright, Patent, Proprietary, etc.) is
essentially a
contract between parties (subjects) regarding the use of resources. The "agreement" is a relationship
between subject and resource. The question of how to best model relationships like this with XACML
attributes is what I would like bring up for discussion. So far two approaches to this problem have

1.) Creating new attribute categories that would represent the relationships. Below is an excerpt
from Hal
regarding this approach.

2.) Determine the relationship at the PIP. This is one approach that IPC profile suggests regarding
of the Agreement-Id, Valid-Agreement-Exists and Number-Of-Valid-Agreements attributes. Below is an
from Erik on some of his thoughts on this topic.

I think Hal is on the right track but hasn't taken it far enough.

I've encountered use cases where I need to reference attributes that are not
part of the subject, but rather are attributes of some entity/object that is
related to subject. For example, attributes of the organization that employs
the subject, or attributes of another person, being the parent/guardian of the
subject. The IPC and EC-US profiles have other examples of a need to reference
attributes of things related to the subject or resource, but not really part
of the subject or resource.

The IPC and EC-US profiles try to shoehorn these attributes of related objects
into the subject and resource categories by "flattening", but this is woefully
inadequate when the relationships are one-to-many or many-to-many.

Consider this simple example as an illustration. I want to indicate that the
subject is an employee of Ajax Inc, a commercial organization, which is a
customer of Widget Inc., a non-profit organization and also the IP owner.
The subject is also a contractor to Widget Inc. The IPC profile would have me
create a request context something like this (from here on I've shortened all
the URIs to just the last part and removed the IncludeInResult XML attributes
in the interest of readability):

<Attributes Category="access-subject">
  <Attribute AttributeId="subject-to-organization-relationship">
    <AttributeValue DataType="anyURI">employee</AttributeValue>
    <AttributeValue DataType="anyURI">contractor</AttributeValue>
  <Attribute AttributeId="organization">
    <AttributeValue DataType="string">Ajax Inc.</AttributeValue>
    <AttributeValue DataType="string">Widget Inc.</AttributeValue>
  <Attribute AttributeId="organization-type">
    <AttributeValue DataType="anyURI">commercial</AttributeValue>
    <AttributeValue DataType="anyURI">non-profit</AttributeValue>
  <Attribute AttributeId="business-context">
    <AttributeValue DataType="anyURI">customer</AttributeValue>
  <!-- Other subject attributes here. -->
<Attributes Category="resource">
  <Attribute AttributeId="ip-owner">
    <AttributeValue DataType="string">Widget Inc.</AttributeValue>
  <!-- Other attributes of the resource. -->

The "flattening" process makes the request context ambiguous and leaves policy
evaluation open to false-positive and false-negative results. Which organization
is the subject an employee of ? Is it both ? Which organization is commercial
and which is non-profit ? There's no way to tell from the request context.

The real world allows objects of many kinds that can be related to each other
in many ways. To capture that reality, XACML needs the ability to represent a
request context that is a graph of objects, rather than a small list of
predefined categories. If I reinterpret a category as an object (i.e., a list
of attributes) and use attribute values of the anyURI data-type as references
between objects, then I can turn the request context into a graph without
changing the syntax at all. Having a predetermined category URI isn't adequate
when there is a need to represent multiple instances of the same kind of object,
as Erik points out. The answer is to decouple the kind of a object from the
category URI by regarding the category URI as simply a unique identifier for
a distinct object. These URIs could be anything, e.g., UUIDs, LDAP URLs,
OID URNs, or whatever. The predefined category URIs become well known aliases
for singling out the particular objects that are significant to the access
attempt, e.g., the subject and resource. These objects are the entry points
into the graph.

If I break out the different entities and relationships into separate
<Attributes> elements and use anyURI values to link them, then I get a
request context that looks like this:

<Attributes Category="access-subject">
  <Attribute AttributeId="relationship-ref">
    <AttributeValue DataType="anyURI">relationship-1</AttributeValue>
    <AttributeValue DataType="anyURI">relationship-2</AttributeValue>
  <!-- Other subject attributes. -->
<Attributes Category="relationship-1">
  <Attribute AttributeId="subject-to-organization-relationship">
    <AttributeValue DataType="anyURI">employee</AttributeValue>
  <Attribute AttributeId="organization-ref">
    <AttributeValue DataType="anyURI">organization-1</AttributeValue>
<Attributes Category="relationship-2">
  <Attribute AttributeId="subject-to-organization-relationship">
    <AttributeValue DataType="anyURI">contractor</AttributeValue>
  <Attribute AttributeId="organization-ref">
    <AttributeValue DataType="anyURI">organization-2</AttributeValue>
<Attributes Category="organization-1">
  <Attribute AttributeId="organization"
    <AttributeValue DataType="string">Ajax Inc.</AttributeValue>
  <Attribute AttributeId="organization-type">
    <AttributeValue DataType="anyURI">commercial</AttributeValue>
  <Attribute AttributeId="business-context-ref">
    <AttributeValue DataType="anyURI">business-context-1</AttributeValue>
  <!-- Other attributes of organization "Ajax Inc." here. -->
<Attributes Category="business-context-1">
  <Attribute AttributeId="business-context">
    <AttributeValue DataType="anyURI">customer</AttributeValue>
  <Attribute AttributeId="organization-ref">
    <AttributeValue DataType="anyURI">organization-2</AttributeValue>
<Attributes Category="organization-2">
  <Attribute AttributeId="organization"
    <AttributeValue DataType="string">Widget Inc.</AttributeValue>
  <!-- Other attributes of organization "Widget Inc." here. -->
<Attributes Category="resource">
  <Attribute AttributeId="ip-owner">
    <AttributeValue DataType="anyURI">organization-2</AttributeValue>
  <!-- Other attributes of the resource. -->

The affiliation of the subject to an organization is represented by an object
(<Attributes> element) containing the subject-to-organization-relationship
attribute. The relationship of one organization to another is represented by
an object containing the business-context attribute. Each of the organizations
is represented by an object containing the organization attribute. Note that
relationship-1, relationship-2, organization-1, organization-2 and
business-context-1 are not predefined URIs. I'm using them as placeholders
for UUIDS, or something similar, that uniquely identify each object.

With the above request context I am able to faithfully represent the situation
I originally expressed in English. I don't expect that the PEP would actually
provide a request context like this. More likely, the context handler would
construct the request context on demand by querying a PIP, given a bare-bones
subject and resource provided by the PEP as a starting point.

The next thing I need to address is how a graph-like request context can be
evaluated by policies.

The first thing to note is that attribute designators use predetermined URIs
for the category, but most of the graph objects have URIs that won't be
known at the time a policy is written, and may vary over time. To overcome
that I need an attribute designator in the form of a function so that I can
feed into it the URI values fetched by other attribute designators.

So an attribute designator like this:


is equivalent to an attribute-designator function:

    <Apply FunctionId="attribute-designator">
      <AttributeValue DataType="anyURI">access-subject</AttributeValue>
      <AttributeValue DataType="anyURI">relationship-ref</AttributeValue>
      <AttributeValue DataType="anyURI">anyURI</AttributeValue>
      <AttributeValue DataType="boolean">false</AttributeValue>

Each of the arguments, especially the first one, is able to be an arbitrary
expression instead of a constant.

The attribute-designator function isn't enough by itself. To do really
interesting things with the graph I find I also need to invoke the iterator
expressions I described on the comment list some time ago:

Armed with the attribute designator function and the ForAny expression I
can apply conditions to any part of the graph. For example, if I want to
test whether the subject is an employee of the IP owner I could use this

<ForAny VariableId="$a">
  <!-- $a is bound to each relationship-ref URI in turn -->
  <Apply FunctionId="and">
    <!-- Return true if and only if the subject-to-organization-relationship of
         the relationship object referenced by $a contains the value "employee". -->
    <Apply FunctionId="anyURI-is-in">
      <AttributeValue DataType="anyURI">employee</AttributeValue>
      <Apply FunctionId="attribute-designator">
        <VariableReference VariableId="$a"/>
        <AttributeValue DataType="anyURI">subject-to-organization-relationship</AttributeValue>
        <AttributeValue DataType="anyURI">anyURI</AttributeValue>
        <AttributeValue DataType="boolean">false</AttributeValue>
    <!-- Return true if and only if the organization referenced by the relationship
         object referenced by $a is the IP owner. -->
    <Apply FunctionId="anyURI-at-least-one-member-of">
      <Apply FunctionId="attribute-designator">
        <VariableReference VariableId="$a"/>
        <AttributeValue DataType="anyURI">organization-ref</AttributeValue>
        <AttributeValue DataType="anyURI">anyURI</AttributeValue>
        <AttributeValue DataType="boolean">false</AttributeValue>

For the record, the answer is false for the graph request context I provided
above. The current profile with "flattening" is incapable of such precision.

I would expect to write up the graph-like request context, the attribute
designator function and iterator expressions as a separate profile since they
are generally useful beyond just the IPC profile.


- Richard


On 2012-10-10 04:04, Hal Lockhart wrote (excerpt):


...I have been aware for some time that the Categories of Subject, Resource, Action & Environment are
insufficient as policies become more complex. The first use case I encountered of this kind was when
to label certain operations as "Administrative". There is a need to associate the attribute not
with the
Resource or Action, but with the combination. For example, all operations on Usage Reports might be
Administrative, whereas only creating an Account might be Administrative, other Account operations
not be.

Some Access Control Models consider Action to be a part of Resource, on the theory that to, for
write a file and to write a database record are not really the same in important ways. I think the
treatment of Action as a first class Category has certain advantages, for example in the case where
to write a policy about ALL reads or writes.

I have opposed using selectors to solve this problem for several reasons. First, it tends to hide
what is
intended in a complicated expression. Second, I think it is undesirable for references to
attributes be
dependent on the means used to obtain them. The foo attribute should be the foo attribute whether it
from LDAP, SAML or SQL. Also it makes the system fragile. If the attribute location changes, the
have to change. I objected to Jan Hermann's proposal that we have Input Message and Output Message
Categories for the same reason. Finally, I consider Selectors to be a mechanism to access
attributes in
request context when a suitable designator has not (yet) been defined. Using a selector to access a
repository seems like a kludge to me.

My preferred alternative in such cases is to define a new category type, say, Resource-Action. A
objection is that we could end up with an explosion of Categories, but while this could be true, I
think it
comes from the real world nature of the problem and thus will be a feature of any solution. The
will be different for every attribute and repository combination, possibly an even larger set.

Notwithstanding all that, I am not sure that is the problem in your case below. I don't think the
patent, or
copyright or license are attributes of the relationship, they are attributes of the Resource
(document), The
relationship is what you are trying to capture in the policy. For example, somebody has licensed
Boeing to
use this document. The policy says that since John Tolbert works for Boeing, he can use the document.

I think the problem in this case is that your attributes are not flat scalars, but contain multiple
I am not prepared to propose a data model, but it seems to me you problem is that the copyright
needs to be qualified by country and perhaps other fields. XACML requires new functions to deal
with new
attribute types as a whole, e.g. GeoXACML, but the existing functions can deal with one field at a



On 2012-10-09, 10:50 PM, Erik Rissanen wrote (excerpt):


...There are different ways to deal with this issue. Defining a new category like you suggest Hal
will not
solve the problem because even if you have a new category, you cannot separate multiple instances
of the
same category, thus you cannot represent multiple relations this way.

My recommendation is to resolve the relations in a PIP. It puts some of the "logic" outside the XACML
but it is the most pragmatic approach, since it does not need any extensions to XACML and leads to
XACML policies.

Using complex data types, either in the form of xml <Content> and selectors, tuples represented as
or XACML data type extensions also work, but mean unwieldy expressions in policies or XACML extensions
need to be implemented in code in the PDP.

One idea I have been toying with in my mind for a long time has been to define a sort of generic tuple
type for XACML, but I have not been able to design a nice and clean set of operators on it, so I have
posted anything on the list. There is a suggestion like this on the XACML comments list, btw.

One could extend on this tuple idea, and when one thinks about the various operations which would be
one ends up re-inventing SQL, so I figured there is little practical value for XACML to do so. This
can be
done on the database side in a PIP. I guess there would be some value in terms of explicit

visibility, but I am not convinced, so I never pursued this thread of thought on the TC list.

Best regards,


To unsubscribe, e-mail: xacml-unsubscribe@lists.oasis-open.org
For additional commands, e-mail: xacml-help@lists.oasis-open.org

To unsubscribe, e-mail: xacml-unsubscribe@lists.oasis-open.org
For additional commands, e-mail: xacml-help@lists.oasis-open.org

[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]