Without attestation, you’re not really doing Confidential Computing.
This post – or the title of this post – has been sitting in my “draft” pile for about two years. I don’t know how this happened, because I’ve been writing about Confidential Computing for three years or so years by now, and attestation is arguably the most important part of the entire subject.
I know I’ve mentioned attestation in passing multiple times, but this article is devoted entirely to it. If you’re interested in Confidential Computing, then you must be interested in attestation, because, without it, you’re not doing Confidential Computing right. Specifically, without attestation, any assurances you may think you have about Confidential Computing are worthless.
Let’s remind ourselves what Confidential Computing is: it’s the protection of applications and data in use by a hardware-based TEE (Trusted Execution Environment). The key benefit that this brings you is isolation from the host running your workload: you can run applications in the public cloud, on premises or in the Edge, and have cryptographic assurances that no one with access to the host system – hypervisor access, kernel access, admin access, even standard hardware access – can tamper with your application. This, specifically, is Type 3 – workload from host – isolation (see my article Isolationism – not a 4 letter word (in the cloud) for more details), and is provided by TEEs such as AMD’s SEV and Intel’s SGX – though not, crucially, by AWS Nitro, which does not provide Confidential Computing capabilities as defined by the Confidential Computing Consortium.
Without attestation, you’re not really doing Confidential Computing. Let’s consider a scenario where you want to deploy an application using Confidential Computing on a public cloud. You ask your CSP (Cloud Service Provider) to deploy it. The CSP does so. Great – your application is now protected: or is it? Well, you have no way to tell, because your CSP could just have taken your application, deployed it in the normal way, and told you that it had deployed it using a TEE. What you need is to take advantage of a capability that TEE chips provide called an attestation measurement to check that a TEE instance was actually launched and that your application was deployed into it. You (or your application) asks the TEE-enabled chip to perform a cryptographically signed measurement of the TEE set-up (which is basically a set of encrypted memory pages). It does so, and that measurement can then be checked to ensure that it has been correctly set up: there’s a way to judge whether you’re actually doing Confidential Computing.
So, who does that checking? Doing a proper cryptographic check of an attestation measurement – the attestation itself – is surprisingly tricky, and, unless you’re an expert in TEEs and Confidential Computing (and one of the points of Confidential Computing is to make is easy for anyone to use these capabilities), then you probably don’t want to be doing it.
Who can perform the validation? Well, one option might be for the validation to be done on the host machine that’s running the TEE. But wait a moment – that makes no sense! You’re trying to isolate yourself from that machine and anyone who has access to it: that’s the whole point of Confidential Computing. You need a remote attestation service – a service running on a different machine which can be trusted to validate the attestation and either halt execution if it fails, or let you know so that you can halt execution.
So who can run that remote attestation service? The obvious – obvious but very, very wrong – answer is the CSP who’s running your workload. Obvious because, well, they presumably run Confidential Computing workloads for lots of people, but wrong because your CSP is part of your threat model. What does this mean? Well, we talked before about “trying to isolate yourself from that machine and anyone who has access to it”, the anyone who has access to it is exactly your CSP. If the reason to be using Confidential Computing is to be able to put workloads in the public cloud even when you can’t fully trust your CSP (for regulatory reasons, for auditing reasons, or just because you need higher levels of assurance than existing cloud computing), then you can’t trust your CSP to provide the remote attestation service. To be entirely clear: if you allow your CSP to do your attestation, you lose the benefits of Confidential Computing.
Attestation – remote attestation – is vital, but if we can’t trust the host or the CSP to do it, what are your options? Well, either you need to do the attestation yourself (which I already noted is surprisingly difficult), or you’re going to need to find a third party to do that. I’ll be discussing the options for this in a future article – keep an eye out.
1 – the TEEs used for Confidential Computing don’t aim to protect against long-term access to the CPU by a skilled malicious actor – but this isn’t a use case that’s relevant to most users.
2 – actually, not that surprising if you’ve done much work with cryptography, cryptographic libraries, system-level programming or interacted with any silicon vendor documentation.