Frequently asked questions¶
What app architecture issues should we consider?¶
We plan to offer more guidance on this in future. Until then consider:
- Dataset size. It must fit in memory in a single CPU machine.
- Slightly lower performance than usual. The extra security checks used by SGX reduce execution performance and the embedded JVM inside the enclave isn't as sophisticated as HotSpot, so peak performance will be lower than normal for Java.
What cryptographic algorithm(s) does Conclave support?¶
The elliptic curve used is Curve25519 for encryption and Ed25519 for signing. The symmetric cipher is AES256/GCM. The hash function is SHA256. The Noise protocol is used to perform Diffie-Hellman key agreement and set up the ciphering keys. As these algorithms represent the state of the art and are widely deployed on the web, there are no plans to support other algorithms or curves at this time.
Which communication channels exist to/from the enclave?¶
The host can exchange local messages with the enclave, and an encryption key is included in the remote attestation. Clients can use this key transparently via Conclave Mail to communicate with the enclave in an encrypted manner.
It's up to the host to route messages to and from the network. Conclave doesn't define any particular network protocol to
use. Mails and serialized
EnclaveInstanceInfo objects are just byte arrays, so you can send them using REST, gRPC,
Corda flows, raw sockets, embed them inside some other protocol, use files etc.
Will using Conclave require any agreements with Intel?¶
Not if you use the latest generation of hardware. Modern chips support a feature called flexible launch control that allows the owner of the hardware to decide which enclaves can be loaded. For example, if you deploy to Microsoft Azure no agreement or interaction with Intel is needed: Azure allows any self-signed enclave.
On older hardware that uses the EPID attestation protocol, to go live you'll need:
- A "service provider ID" that grants access to the Intel attestation servers. This is free.
- To sign a (very simple, one page) agreement that you won't sign malware. Intel will whitelist your signing key and Release Mode then becomes available. This is also free.
Due to Conclave's design enclave clients don't need to interact with Intel at any point. The
host does it and then publishes to clients a serialised
EnclaveInstanceInfo object. The Conclave client libraries
embed the necessary certificates to verify Intel's signature over this data, and the integrity of the object
is checked automatically when it's deserialized. The developer doesn't have to do anything: it's all fully automated.
How are upgrades performed when patches are released?¶
There are several components which might need an upgrade when a vulnerability is found. It might involve updates to the Intel platform services software (a daemon and some libraries on the host Linux machine), the BIOS or UEFI versions and configuration, Conclave, or even the application code itself.
Some of these updates might involve rebooting the system, which can cause downtime depending on the security requirements of each application. Conclave Mail is designed to be an asynchronous messaging system, so restarts to upgrade to new versions would be seen by clients as a latency spike rather than downtime.
It might be possible to perform a rolling update, permitting applications to continue working on a potentially compromised system while the rollout is being performed. Stateless enclaves are easier to interchange, since no state needs to be transferred to other enclave instances. In the case of stateful enclaves, the possibility for doing state transfers depend on the application requirements, since data sealed on a system cannot be unsealed on a different system.
Who and what will we be trusting exactly?¶
You must trust the CPU vendor. However, you have to trust your CPU to use computers in the first place, regardless of whether you're using SGX or not. It's well within the capabilities of CPU microcode to detect code patterns and create back doors or rewrite programs on the fly. A few minutes reflection will demonstrate that you would usually have no way to detect this, unless you can recruit some other CPU that isn't back doored to assist.
You don't need to trust the owner of the hardware on which you run.
We will allow audits of Conclave's source code to paying customers, so the full code inside the enclave can be verified directly. Therefore, you don't have to trust R3 if you have a commercial relationship with us.
The users of an enclave-backed service must either verify the source code of the enclave themselves, to ensure it’s "behaving properly" (i.e. not sending secrets to the host), or they must outsource this task to a third party auditor they trust to accurately summarise the enclave's behaviour for them.
Verifying the enclave measurement via remote attestation ensures the source code matches the actual code running in the enclave. This also lets you check the remote system is fully up to date, and that the enclave is not running in debug mode (i.e. that the memory really is encrypted).
How long should we trust an attestation verification report?¶
When a flaw in the system is found Intel performs a "trusted computing base recovery" operation. This involves releasing security updates and adjusting their servers to start reporting that old versions are known to be compromised. Thus you shouldn't accept a remote attestation for too long.
Intel provides official guidance with some recommendations:
- For organizations with low risk tolerance (for example, banking), the attestation frequency policy might be once per day or once per week.
- Organizations with higher risk tolerance may define a frequency policy of once every 180 days.
- A typical frequency policy is once every 30 days.
At the moment the developer is responsible for writing code that restarts the enclave from time to time. Restarting
the enclave will force a re-attestation and thus make the
EnclaveInstanceInfo fresh again. How often this happens
should be synchronised between the host and client side tool (but remember to make the host server refresh more
frequently than the client requires, to avoid synchronisation errors).
Isn't putting a whole JVM into an enclave a bit fat?¶
The so-called native images produced by Conclave don't include a traditional full JVM like HotSpot. All code is compiled ahead of time and dynamic classloading isn't possible, which gives a model more similar to that of Rust. The runtime aspects of the enclave are the garbage collector (which is certainly worth its weight in gold when lack of memory safety bugs is considered), and some small pieces of code for miscellaneous runtime tasks like the generation of stack traces.
Keeping an enclave small is a good idea to reduce the chance of security critical bugs, but this logic mostly applies when writing code in unsafe languages like C, C++ or when using unsafe blocks in Rust. The runtime in Conclave enclaves exists mostly to prevent security bugs in the rest of the code, additionally, the runtime is itself written in Java meaning the same bounds checking and memory safety technologies that protect your code also protect the runtime. The lack of dynamic code loading (when using normal JVM bytecode) means even the runtime functions aren't exposed to attackers, so on balance we feel this is a clear security win.
What about side-channel attacks?¶
Please see the discussion of side channel attacks in the security documentation.
Development and deployment¶
What Kotlin/Java versions are usable?¶
Inside the enclave you can use Java 8 or 11, and Kotlin 1.4.
At which points in time is a connection with Intel needed?¶
When using DCAP attestation (which requires the latest hardware), e.g. on Microsoft Azure, no direct connection to Intel is required. The cloud provider runs caching proxies in-cloud and Conclave uses them automatically.
When using the older EPID attestation protocol a connection to Intel's servers will be made by the
aesmd daemon and
the host program when an enclave is loaded, as part of the startup sequence. See the information about
machine setup to learn about firewall and proxy configuration.
Can I print debug output to the console from my enclave?¶
Conclave enclaves built for debug and simulation support output to the console from inside the enclave
through the use of
System.out.println(). Release builds of enclaves do not support printing to the console.
System.out.println() in release builds of enclaves is allowed but the output is discarded
inside the enclave. This is to prevent accidental leakage of enclave state through the use of debug logging.
Can I load more than one enclave at once?¶
Yes, but each enclave must be a separate (Gradle) module. One module can only have one enclave entrypoint. On current hardware you may need to keep the enclaves small in order to avoid running out of fast EPC RAM, in which case the enclave memory will be 'swapped' in and out of EPC, at a considerable performance penalty.
You can consider an alternative approach. Because of how Conclave Mail is designed, enclaves can message each other when offline. Therefore, it's possible to load one enclave, have it send a message to another enclave, unload it, load the second enclave and have it process the message from the first. Splitting up app logic in this way can be useful for allowing modules to be upgraded and audited independently.
Why does my enclave build fail?¶
We use the GraalVM native image technology. Builds can fail if you are trying to dynamically load bytecode (this doesn't work yet), if you use a feature not supported by our version of Native Image, or if you're on Windows/macOS and you don't allocate enough RAM to Docker. The build is memory intensive, so please allocate at least 6GB of RAM to Docker and possibly more. However you don't need to do builds all the time when working on an enclave. You always have the option of using mock mode to directly load the enclave code into the host JVM, which gives a regular Java development experience.
Java object serialization doesn't work?¶
On GraalVM native image builds, this is because Java object serialization isn't implemented yet. It will arrive in
future versions. On Avian builds this is because Java object serialization is configured with a filter that blocks all
usage by default. This is to protect the enclave against "mad gadget" attacks.
You can make it work again by setting a less restrictive serialization filter on your
consider some other serialization mechanism.
How do I control the Native Image parameters?¶
You can add flags to the
native-image command line when building to optimise your enclave binary and control the
details of how it's compiled. Create a file called
in your enclave module and set it to contain:
This doesn't let you alter or override the flags we already specify in order to create a working enclave.
Do you have plans to support AMD SEV?¶
Not at this time. We will re-evaluate future versions of SEV. The main problems are:
- AMD SEV is very strongly oriented around the protection of VMs in the cloud. Every protected domain has an owner that has full access to it. This makes it useless for multi-party computations where nobody should have access to the full state of the calculation.
- It has no equivalent of SGX TCB recovery, meaning flaws permanently break the system. Prior versions of SEV have been rendered useless by the discovery of numerous fatal bugs in AMD's firmware. Although patches were made available there was no way to remotely detect if they are actually applied, which made patching meaningless.
Additionally, in SEV remote attestation is randomized which means you can’t ask a remote host "what are you running". You are expected to know this already (because you set up the remote VM to begin with). This doesn't fit well with most obvious enclave APIs.
AMD and Intel are in increasingly strong competition, and so we expect AMD to catch up with the SGX feature set in future. Intel are also working on an SEV-equivalent feature for protecting entire (Linux) virtual machines from the host hardware. Conclave may add support for this at some point in the future as well.
Do you have plans to support ARM TrustZone?¶
Not at this time. ARM TrustZone doesn't have any form of remote attestation support. It's meant for hardening mobile phone operating systems and supported use cases don't really go beyond that. ARM will likely enhance their CPUs to support remote attestation in future, and we will re-evaluate when such support ships.
Do you have plans to support AWS/Nitro?¶
Nitro is the name Amazon uses to refer to in-house security technology on their custom servers that restricts access by AWS employees. It's not an enclave and Amazon must still be assumed to have access to all your data, institutionally, because they design and implement Nitro, thus you have only their assurance that there are no back doors or internally known weaknesses.
As it's not an enclave we currently have no plans to support Nitro.
Do you have plans to support cloud providers as the root of trust?¶
Although Nitro is not an enclave, some cloud providers do allow you to generate an attestation-like structure that asserts what disk image was used to boot a virtual machine. For cases where you're willing to trust a cloud provider but still want an auditable code module as part of your app, we may consider adding support for this in future.
Why are you using Intel SGX versus mathematical techniques?¶
Mathematically based secure multi-party computation and zero knowledge proofs are closely related and have similar tradeoffs/problems such as high complexity, low performance, difficult to use without intense dedicated training and not stable (i.e. the best known algorithms change frequently).
SGX is relatively simple, high performance, much easier to use, and stable over time.
However, we acknowledge the limits of relying on hardware based solutions and are interested in bringing easy to use circuit-based cryptographic algorithms to market one day.