Let's say we want to force all of our client to use pairwise subs, but a single "merchant" needs to implement several clients where subs should remain the same for all those clients.

Merchant A
- client x
- client y
- client z

 Merchant B
- client m
- client n

You can assume the sector_identifiers are the same across all clients owned by a merchant

It should not be possible to correlate activities between Customer A and Customer B (at least not from their side). They should, however, be able to correlate user activities between their own clients.

Which implementation of pairwise subs is better suited for supporting this scenario? I'm leaning towards the protocol mapper solution. It should be easier to create custom mappers with merchant-wide configuration (e.g salts).

On Mon, 22 Aug 2016 at 22:40 Marek Posolda <mposolda@redhat.com> wrote:
The question is, should we really introduce another SPI for that? Doesn't it mean uneccessary complexity? Also add the new options directly to the client for the feature, which is likely interesting just for quite limited amount of people?

IMO it's fine if this is implemented as protocolMapper?

Few thoughts:
- We can have abstract superclass like AbstractPairwiseSubGeneratorMapper . The subclasses will just need to implement method "generateSub" . We can have just one concrete impl, which will use SHA-256( sector_identifier || local_sub || salt )

- The sector_identifier_uri will be a configuration option of this protocolMapper implementation.

- If protocolMapper is not added to client, the client will just use the public subjects. By default it's not added, which ensures backwards compatibility and public subjects by default. Note that with this approach, we don't even need subject_type option on the client.

- The salt can be generated lazily at the first time when mapper is used.

- The validation can be done at the moment, when protocolMapper is going to be created/updated. Right now, we don't have validation callback during protocolMapper creation/update. However Bill is going to add support for that into generic componentModel. So if we're going to refactor protocolMapper to use the new generic component model, we will have validation callback available to protocolMapper SPI. The validation will fail if array of redirect_uri from sector_identifier_uri doesn't contain the uris from redirect_uri of particular client (including wildcards etc).

- If client is updated and it's redirect_uri are changed, we won't be able to catch this, however this is not strictly required per specs per my understanding. At least the dynamic client registration specs sais [1]

"The values registered in redirect_uris MUST be included in the elements of the array, or registration MUST fail. This MUST be validated at registration time; there is no requirement for the OP to retain the contents of this JSON file or to retrieve or revalidate its contents in the future. "

[1] http://openid.net/specs/openid-connect-registration-1_0.html#SectorIdentifierValidation


Marek


On 22/08/16 15:50, Martin Hardselius wrote:
Ok, thanks for the clarification.

Where would it make sense to put the PairwiseSubGeneratorSpi? Which package, that is?

On Mon, 22 Aug 2016 at 14:51 Stian Thorgersen <sthorger@redhat.com> wrote:
On 22 August 2016 at 14:16, Martin Hardselius <martin.hardselius@gmail.com> wrote:
"IMO it's sufficient to document the algorithm to create the sub, which should make it clear that the origin in the redirect uri will affect the sub."

Reasonable. :)

"One client could also have multiple redirect uris with different origins so could get different sub's generated depending on the redirect uri used."

That case is probably caught by
[...] If there are multiple hostnames in the registered redirect_uris, the Client MUST register a sector_identifier_uri[...]

Yes, but I meant from a documentation perspective. It should be clear from the documentation that is the case.
 

On Mon, 22 Aug 2016 at 10:42 Stian Thorgersen <sthorger@redhat.com> wrote:
IMO it's sufficient to document the algorithm to create the sub, which should make it clear that the origin in the redirect uri will affect the sub. One client could also have multiple redirect uris with different origins so could get different sub's generated depending on the redirect uri used. 

On 22 August 2016 at 09:58, Martin Hardselius <martin.hardselius@gmail.com> wrote:
Sounds fair enough.

What about the case where you don't provide a sector_identifier_uri, set up a single redirect uri on myhost and then later go on to change that redirect uri to something on myotherhost? That would change the sector_identifier and subsequently all the user subs. Do we protect against such "mistakes" or do we warn people (in the docs and/or admin gui) that it's not a good idea to do that?

On Mon, 22 Aug 2016 at 09:38 Stian Thorgersen <sthorger@redhat.com> wrote:
We need to follow the spec and verify that sector_identifier_uri points to a URL that contains the corresponding URIs. As an enhancement we could support wildcards in the JSON returned by sector_identifier_uri. For example if it returns:


A client with the redirect uri 'https://www.myotherdomain.com/myapp' would work

On 18 August 2016 at 15:09, Martin Hardselius <martin.hardselius@gmail.com> wrote:
Speaking of subject_identifier_uri

From the spec:

"When a sector_identifier_uri is provided, the host component of that URL is used as the Sector Identifier for the pairwise identifier calculation. The value of the sector_identifier_uri MUST be a URL using the https scheme that points to a JSON file containing an array of redirect_uri values. The values of the registered redirect_uris MUST be included in the elements of the array."

What's your stance on sanity/health checking the sector_identifier_uri? URI validation is one thing, but should we also make a request to the uri in order to validate the response?

The spec also mentions that the sector_identifier_uri isn't strictly required if a client has all it's redirect_uris under the same domain. How do we deal with changes to clients if the sector_identifier_uri isn't required for enabling pairwise subs?

Example:

I create a client, enabling pairwise subs. Valid redirect_uris are [ https://www.mydomain.com/* ]. The sector_identifier would be mydomain.

Later on, I update the redirect_uris to [ https://www.mydomain.com/*, https://www.myotherdomain.com/* ] Do we

* keep the old sector_identifier, or
* reject the update, or
* allow the update if a valid subject_identifier_uri is provided at mydomain, or
* just allow it and let the client devs deal with the consequences, or
* take some other path you can think of ?

Having the sector_identifier_uri as a hard requirement seems safer, but it's also means more work implementing a client. Leaving it optional is a lot more scary.


On Thu, 18 Aug 2016 at 10:45 Stian Thorgersen <sthorger@redhat.com> wrote:
Makes sense to make this pluggable. The default should probably SHA-256( sector_identifier || local_sub || salt ).

We would love a PR for this, but there's a few bits required:

* OIDC clients need option for subject type and sector_identifier_uri. If not set it should default to public so existing clients continue to work. These options can just be set as client attributes so there's no need to update db schema
* Admin console update for the above
PairwiseSubGeneratorSpi and default implementation
* Generation of salt for clients. This should be done when a client sets subject type to pairwise
* Tests and docs

I'd say the PairwiseSubGeneratorSpi signature should probably be:

* public String getPairwiseSub(UserModel user, ClientModel client)

It might even be an option to let the default PairwiseSubGenerator provider create the salt and store it as an attribute on the client, making that part pluggable as well.

On 17 August 2016 at 15:59, Martin Hardselius <martin.hardselius@gmail.com> wrote:
I'm going to bump this, as I want to continue the discussion/provide some input.

Does it make sense to support more than type of pairwise subject identifier generator? E.g through a PairwiseSubGeneratorSpi?

Let's say I want to generate the pairwise sub as a salted hash: sub = SHA-256( sector_identifier || local_sub || salt )
To me, it makes sense to allow for per-client salts. These salts should probably be generated and persisted during client creation. Thoughts?

On Fri, 12 Aug 2016 at 09:57 Martin Hardselius <martin.hardselius@gmail.com> wrote:
Thank you for your response. Did not see that ticket before. Great news!

I looked into using protocol mappers to achieve this, and while it would work I'm worried that once KEYCLOAK-3422 has been resolved and included in a proper release we would run into migration issues if the method used for calculating "native" pairwise subs is different from what we implement. Clients could loose / be forced to re-register all their users if we decide to switch. The example methods in the spec are just that. Examples. Maybe the method/alg for computing the pairwise sub should be pluggable?

-- 
Martin

On Thu, 11 Aug 2016 at 17:15 Marek Posolda <mposolda@redhat.com> wrote:
Sorry for late response.

We have JIRA created for that. You can possibly add yourself as a watcher. See https://issues.jboss.org/browse/KEYCLOAK-3422

Maybe an alternative for you is to use protocolMappers. That should allow you to "construct" the token for particular client exactly how you want and also use the different value for "sub" claim.

Another possibility is, to handle this on adapter side. We already have an adapter option "principal-attribute", which specifies that application will see the different attribute instead of "sub" as subject. For example when in appllication you call "httpServletRequest.getRemoteUser()" it will return "john" instead of "123456-unique-johns-uuid" . See https://keycloak.gitbooks.io/securing-client-applications-guide/content/v/2.1/topics/oidc/java/java-adapter-config.html

Hopefully some of the options can be useful for you?

Marek


On 02/08/16 14:13, Martin Hardselius wrote:
Me and my team are working towards getting Keycloak, customized for our needs, into production but we've identified the need for Pairwise Subject Identifiers as we don't want to expose internal user ids.

Right now, the only subject_types_supported seems to be "public". Are there any near-future plans to include "pairwise"? Can we pitch in with a PR to make this happen as soon as possible?

Links to relevant sections in the spec:

http://openid.net/specs/openid-connect-core-1_0.html#SubjectIDTypes

-- 
Martin


_______________________________________________
keycloak-dev mailing list
keycloak-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-dev


_______________________________________________
keycloak-dev mailing list
keycloak-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-dev