Re: [Openstack] Plans for Trusted Computing in OpenStack
The assessment criteria is clear, that IaaS provider measures the whole cloud stack every time when it creates a new release, and then is compared to the run-time measurements on the running stack on the hosts in so-called attestation process. A failed attestation implicates the host is untrusted because the stack has been changed. This is because IaaS provider builds and trusts its own deployed stack. So the remote attestation service is built to help IaaS provider check whether the trusted stack has been changed on the hosts. Here the trust criteria is clear; However, an external CA doesn't build the stack, and I don't know how CA can judge whether a specific cloud stack is trusted or not, even when IaaS provider is asked to share. A malicious IaaS provider can provide a evil stack with holes on any point. Here the 'trust' criteria is not clear. BTW, I do think your proposal might be a good complement to the existing remote attestation service from another angle. It's possible that remote attestation framework, or scheduler, or other involved components contains bug, which lead to VM running on untrusted host even when the attestation fails. In that case, give the VM capability to detect its own secret sounds like an acknowledgement to a successful attestation, since a failed attestation can never inject the secret. :) Thanks Kevin From: Nicolae Paladi [mailto:n.pal...@gmail.com] Sent: Thursday, November 15, 2012 12:19 AM To: Tian, Kevin Cc: Dugger, Donald D; openstack; Li, Susie; Wei, Gang; Maliszewski, Richard L Subject: Re: [Openstack] Plans for Trusted Computing in OpenStack That is correct, the variety of versions, components and patches is the first thing that comes to everyone's mind with this approach. But the idea is not to have a trusted third party/CA that would be able to assess _all_ combinations. With both approaches, the 'assessment' is left out as a stub or assumption (whichever you prefer). In this case it doesn't actually matter who does the assessment - the IaaS provider or a CA, since the assessment criteria are the unsolved issue. The use case we're examining here is when a certain IaaS provider is contracted to supply IaaS to a client with special requirements, let's call it C. That would mean, e.g: 1. Potentially not all hosts would be used to deploy VMs for C 2. Potentially patches and versioning might go through a separate upgrade flow for those hosts To summarize and address your concerns here: * imo once a client is concerned enough to require trusted hosts, the use of using external assessment case becomes valid (and assessment by the IaaS provider becomes useless) * wrt to variation of the software stack, the big issue is the assessment criteria rather than the assessing party. Cheers, /Nico. On 14 November 2012 04:27, Tian, Kevin kevin.t...@intel.commailto:kevin.t...@intel.com wrote: I would see the major blocking issue on this: However, imo an IaaS provider that claims to offer trusted hosts but hesitates to reveal the software stack of it's hosts to an external auditor (CA in this case) would have issues with credibility every IaaS provide has its own software stack, picking different components, with different versions, possibly adding different patches. The stack can be changed frequently. Rackspace says they can publish a build in less than 1hrs to deploy a new stack. Having an external CA to hold credentials for such large uncertain combos is almost a mission impossible. :) Thanks Kevin From: Nicolae Paladi [mailto:n.pal...@gmail.commailto:n.pal...@gmail.com] Sent: Tuesday, November 13, 2012 11:46 PM To: Dugger, Donald D Cc: openstack; Tian, Kevin; Li, Susie; Wei, Gang; Maliszewski, Richard L Subject: Re: [Openstack] Plans for Trusted Computing in OpenStack Hi, I agree that the use case of a trusted IaaS provider (with possibly compromised nodes) is a valid one and should have support in the openstack codebase, although it seems rather dicey to trust the IaaS provider which does not trust it's own hosts. And your understanding is correct, the idea is to add a 3rd party 'CA' with the aim to assess the integrity of the hosts based on the data produced by the TPM. What I am advocating here is the scenario where the IaaS can not be trusted. In this case the CA would only gain information about the software stack of the IaaS provider's hosts, necessary to perform the attestation. However, imo an IaaS provider that claims to offer trusted hosts but hesitates to reveal the software stack of it's hosts to an external auditor (CA in this case) would have issues with credibility. Wrt complexity, this would require: * sending the attestation information externally to the 'CA' and taking a launch/not launch decision based on the result of the attestation. Even if the untrusted IaaS launches the VM, the client can easily detect the fraud. * on the compute host, decrypting the nonce provided by the client
Re: [Openstack] Plans for Trusted Computing in OpenStack
I would see the major blocking issue on this: However, imo an IaaS provider that claims to offer trusted hosts but hesitates to reveal the software stack of it's hosts to an external auditor (CA in this case) would have issues with credibility every IaaS provide has its own software stack, picking different components, with different versions, possibly adding different patches. The stack can be changed frequently. Rackspace says they can publish a build in less than 1hrs to deploy a new stack. Having an external CA to hold credentials for such large uncertain combos is almost a mission impossible. :) Thanks Kevin From: Nicolae Paladi [mailto:n.pal...@gmail.com] Sent: Tuesday, November 13, 2012 11:46 PM To: Dugger, Donald D Cc: openstack; Tian, Kevin; Li, Susie; Wei, Gang; Maliszewski, Richard L Subject: Re: [Openstack] Plans for Trusted Computing in OpenStack Hi, I agree that the use case of a trusted IaaS provider (with possibly compromised nodes) is a valid one and should have support in the openstack codebase, although it seems rather dicey to trust the IaaS provider which does not trust it's own hosts. And your understanding is correct, the idea is to add a 3rd party 'CA' with the aim to assess the integrity of the hosts based on the data produced by the TPM. What I am advocating here is the scenario where the IaaS can not be trusted. In this case the CA would only gain information about the software stack of the IaaS provider's hosts, necessary to perform the attestation. However, imo an IaaS provider that claims to offer trusted hosts but hesitates to reveal the software stack of it's hosts to an external auditor (CA in this case) would have issues with credibility. Wrt complexity, this would require: * sending the attestation information externally to the 'CA' and taking a launch/not launch decision based on the result of the attestation. Even if the untrusted IaaS launches the VM, the client can easily detect the fraud. * on the compute host, decrypting the nonce provided by the client (and _sealed_ by the CA to the trusted configuration of the host). That will add up to the codebase but is rather trivial (involves mostly interacting with the TPM). Now, there are some design choices to be made, e.g. whether the host communicates its attestation credentials directly to the CA in an https session or the scheduler does that. However, it does not change the important point that the client can verify that the VM was started on a trusted host, without having to rely on the IaaS provider. A common topic to both models (trusted or untrusted IaaS provider) is the question of security profiles for the hosts -- are you considering binary values (trusted/untrusted) or some finer-grained scale? Happy to hear your opinions and continue the discussion; cheers, /Nico. On 12 November 2012 21:23, Dugger, Donald D donald.d.dug...@intel.commailto:donald.d.dug...@intel.com wrote: Nicolae- We've been working under the assumption you have trust the IaaS provider (individual nodes might have been compromised somehow but you trust the provider itself). I think what you are looking at is adding a 3rd party CA which is significantly increasing the complexity of the solution and potentially exposing the IaaS's infrastructure to a 3rd party, probably not desirable to the IaaS provider. I've added some others to the thread who can chime in with their opinions. -- Don Dugger Censeo Toto nos in Kansa esse decisse. - D. Gale Ph: 303/443-3786tel:303%2F443-3786 From: Nicolae Paladi [mailto:n.pal...@gmail.commailto:n.pal...@gmail.com] Sent: Wednesday, November 07, 2012 7:42 AM To: Dugger, Donald D Cc: openstack Subject: Re: [Openstack] Plans for Trusted Computing in OpenStack Hi, so basically my questions/thoughts about support for TC in OpenStack are based on a somewhat different attack model where the IaaS is actually not trusted. That is in contrast with the Trusted Compute Pools, where the scheduler/trusted_filter is assumed to reject the host as a candidate for running the VM if it does not have a corresponding trust value. However, nothing prevents a really evil IaaS deployment to ignore this trust value and go ahead, launch the VM and return it to the client. So there's an improvement suggestion focusing on that part. The model that I have in mind assumes both no trust in the IaaS setup/provider. So the gist is that: 1. Client could upload a secret encrypted with the public key of the authentication service (possible to include in the extra_specs) 2. The Attestation Service, after verifying the compute host could bind the secret to the hosts trusted configuration, so that the host can inject the secret into the VM With this approach, a malicious IaaS provider can still launch the VM on an untrusted host, but now he client can verify that the VM has been started on a 'trusted' host. So the questions around this are -- 1. Is the scenario of an untrusted IaaS
