Re: [Lurk] [E] Re: keeping private key private

[[e-mail]]

Hi Daniel – Good to see clarification below to at least some of the key questions that had come up at the BoF re LURK protocol.  And keeping with clarifications, there was at least one more conversation that I recall where question were raised around changes to cipher suites needed for clients to support LURK. Has there been any progress in LURK as to whether those changes can be quantified and  scoped?

It would also be good to see if the discussions do not almost exclusively limit itself on the certs, as it appears, there are choices, namely, the session key interface; short term certificates; delegated credentials etc. So, importantly, if there is consensus that discussion is merited on content and separation of keys from the content, that would then help to focus discussion on above choices. In terms of use case, there  is/was consensus on at least the CDN use case and would be good to discussion also moves towards other possible use cases.

Best
Sanjay

        [cid:[e-mail]545870]



From: Lurk [mailto:[e-mail]] On Behalf Of Daniel Migault
Sent: Friday, October 14, 2016 2:52 PM
To: Daniel Kahn Gillmor
Cc: [e-mail]
Subject: [E] Re: [Lurk] keeping private key private

Hi Daniel,

Thank you very much for the feed backs. I will try to clarify the problem we are solving, expose why LURK in my opinion address this problem in a "wise" way and expose why short term certificate (STC) does not solve this problem.

Note that I am not opposed to STC, I think this is something good to be done, however, it is complementary to LURK and not an alternative to.
The mail may be long, but feel free to split it into different threads. I will be more than happy to clarify any aspects.

I agree with you that information delivered by a web site is what we must be focused on and this is actually the goal of LURK* (in a TLS context). Although the secret key material (SKM) and the information of a web site are two different assets, they are not independent and protecting the information requires the protection of the SKM. Information is characterized at least by a content and an origin. An un-authenticated content such as a software for example, does not have the same value if it is being downloaded from the editing company or from public content sharing site. The motivation of LURK is to provide the user the means to trust the origin and as such the information delivered by the web site.

As a result, SKM must be protected as long as the information is delivered, independently of the SKM characteristics which includes its life time.

I also clearly agree with you that protection of the SKM should be performed with limited cost and as simple as possible.

It seems to me that LURK meet these criteria, which is the reason we would like it to be standardized. As an observation, LURK like solution have been deployed by CDN for example, and I believe their goals are not to consider the most expensive and most complex solution. That said they might be other and better solutions.
There has been some discussions on whether short term certificates (STC) address the problem we want to solve that is either preventing the leak of the SKM or in a TLS context providing means to trust the origin of the information. In addition, there were also discussions on whether STC are also less expensive and more salable than LURK. I disagree that STC is an alternative to LURK.

    a) First STC does not address problem we are trying to solve, i.e "ensuring the origin of the information". In that context, STC limits the use of a SKM that leaks to a short period of time. The improvement over long term certificate is only valid if leakage of the SKM only happen accidentally - that is one time or when retrieving the SKM takes significantly longer than the life time of the SKM. Such limitation to the protections of the SKM are ways too narrow. SKM does not address for example any attacks that can provide access to the SKM in an timely manner. Such attacks simply needs to be re-done after any SKM renewal. Such attacks includes for example privilege escalation vulnerabilities which there is little ways to insure they will not happen. Heartbleed is typically an existing attacks that only took an RTT to retrieve the SKM. STC does not either address this attack - which we know exists - and may take years to be disclosed. I see LURK ways more powerful as it address the root cause of the problem instead of limiting the consequences of a leak. .

    b) Second, STC are really not inexpensive at all.
        b. 1) I might agree that generating a key is inexpensive. However we should also consider that generating keys with a high randomness frequently might not be as inexpensive as expected.
        b .2) That said certificates are not limited to key generation and includes signing by the CA as well as revocation when needed. Even though the cost of signing by the CA decreases, if the number of signing operations increases in several order of magnitude, the overall cost may not decrease. Especially the cost may remain significantly high for EV certificates and for these certificates STC may happen to be very expensive.
        b. 3) STC as any certificate based solution does not provide appropriated delegation mechanisms. As result, in a context of CDN, any time the content owner changes the CDN a complex revocation procedure might be required. Complexity may be due to the use of EV certificates as well as the interrelations between the CDN and the content owner.

    c) Third, STC as certificate based solution considerably weaken the security.
        c. 1) As certificates does not provide delegation mechanisms, the delegation to a CDN often result in providing the SKM to a thrid party. This goes against the principles of public cryptography, resulting in a authentication that provides little benefit over self signed certificates.
        c. 2) EV may present delays that are non compatible with STC, as a result, the use of STC in our context may be limited to DV certificates which is the weakest certification level.

LURK instead
    a) solves the problem we are trying to solve as it prevents the leakage of the key.
    b) is inexpensive as it does not affect the operations performed by the CA. At least the cost remains the same as it is today. I think one reason people see this solution as expensive is that most people think of the key server as an HSM. HSM is probably an expensive alternative, but there also less expensive alternatives that includes the use of open hardware security modules [1], software implementation of the key server in which case the cost is really the instantiation of the VM. This is typically less expensive than the resulting operations, traffic from STC.
    c) is secure and with multiple level of security. The key server may be only hosted by the content owner, provided to CDNs in different ways. Note that providing an obfuscated key server (like an HSM) to the CDN still provide more security than providing the key and limit the risk of leakage. In addition by clearly splitting the SKM and session termination functions, LURK  is likely to ease the deployment of TLS over CDNs and as result the deployment of TLS.

I agree with you that having a key server adds an exchange between the server terminating the TLS session and the key server which may result in higher latency. The latency clearly depends on where the key server is located. In case a high latency matters, the key server may be deployed close to the terminating nodes. Of course there is a trade off between high security requirements and low latency requirements. LURK makes this trade off possible whereas certificates clearly prioritize the latency over security. That said, the acceptable latency also depends on the information delivered by the web site. Suppose the information is an update, latency does not really matter, similarly if the content is a movie a additional .5 second might be acceptable too.

I do not really agree that LURK presents a scalability issue. I agree that the model of LURK is more centralized that the model proposed by SFT, similarly to what DNS is versus hostnames files. Of course the key server needs to be deployed so that it serves the expected traffic, and a single instance of key server may not be appropriated. On the other hand, by centralizing the private key operations, it makes the deployment of a new private key way easier. Only the key servers have to be provisioned with the private key. Maybe we could also associate the key server with a distribution of certificates which STC may designed. Again I do not see STF are competing LURK but more complementing it.


BR,
Daniel


* For clarification let's call LURK the solution that consists in remotely perform secret key operations.
[1] https://www.crowdsupply.com/cryptech/open-hardware-security-module

On Wed, Oct 12, 2016 at 5:45 PM, Daniel Kahn Gillmor <[e-mail]<mailto:[e-mail]>> wrote:
Hi Daniel--

On Wed 2016-10-12 13:10:04 -0400, Daniel Migault wrote:

> Currently delegating a TLS service to a third party or multiple edge nodes
> is mostly based on distributing the full private key to multiple edge nodes
> that may even be in a different domain. Such a distribution prevents the
> key owner to keep the control of the private key as well as to prevent the
> key to leak. For example, with LURK, the Heartbleed attack would have had
> very limited impact as the private key would not have been hosted on the
> edge server. By design, LURK would have prevented the key to leak.

What's more important -- the secret key material, or the traffic and
information on your website that it protects?

When the certificate is long-lived, there are good arguments for the
secret key being more valuable, because it can be misused in the future
when we don't yet know what the value of of the traffic and information
will be.

It's also conceivable that a single cert (e.g. *.example.biz<http://example.biz>) could be
more valuable than any one site it applies to.

But when the certificate is short-lived and narrowly-scoped, the things
the cert is trying to protect are proportionally more valuable, though,
right?

secret keys are cheap!  certificates are (increasingly) cheap!  Making
cheap things less powerful and using them freely seems like a much
simpler approach than trying to invest certificates with lots of power
and then trying to find ways to avoid the misuse of that power.

> My understanding is that there are currently no mechanisms that prevent the
> keys to leak in the case of TLS. There is a huge demand from the industry
> to provide mechanisms that protects the private key from leakage. The
> reason we focused on TLS is that by limiting the use of a single protocol
> it provides more control on the usage of the key. However, the protection
> of the private key can also be handled in a more generic way outside the
> scope of TLS.  I would like to understand whether mechanisms to prevent
> leakage of the private key should not be handled in the context of TLS and
> instead being handled in a more generic way.

we have (a few) generic mechanisms already.  PKCS#11, as horrible as it
is, provides for secret key isolation.  "PKCS#11-RPC" was floated as one
idea in the LURK BoF, iirc, but i think it was (mostly) dismissed,
though i don't remember why.

But remote use of secret key material is also high-latency and prone to
bottlenecks -- are these engineering costs worth the benefits?

some people might amortize the costs of the high-latency handshake by
establishing and encouraging the use of TLS sessions; this is in effect
asking users to remember short-lived certs -- those sessions are going
to still be valid authenticators (to those end users) even after the
keyholder decides to rescind her grant of use of her secret key.

So there are a set of tradeoffs involved, and it's not clear to me that
the thing you're trying to protect is worth the energy and cost that
would be invested here.  can you explain how you see those tradeoffs?

All the best,

      --dkg

_______________________________________________
Lurk mailing list
[e-mail]<mailto:[e-mail]>
https://www.ietf.org/mailman/listinfo/lurk