tpm.dev.tutorials/Attestation/README.md

What Attestation is

A computer can use a TPM to demonstrate:

• possession of a valid TPM

• it being in a trusted state by dint of having executed trusted code to get to that state

• possession of objects such as asymmetric keypairs being resident on the TPM (objects that might be used in the attestation protocol)

Possible outputs of succesful attestation:

• authorize client to join its network

• delivery of configuration metadata to the client

• unlocking of storage / filesystems on the client

• delivery of various secrets, such credentials for various authentication systems:

  • issuance of X.509 certificate(s) for TPM-resident attestaion public keys

    For servers these certificates would have dNSName subject alternative names (SANs).

    For a user device such a certificate might have a subject name and/or SANs identifying the user or device.

  • issuance of non-PKIX certificates (e.g., OpenSSH-style certificates)

  • issuance of Kerberos host-based service principal long-term keys ("keytabs")

  • service account tokens

  • etc.

• client state tracking

• etc.

Possible outputs of unsuccessful attestation:

• alerting

• diagnostics (e.g., which PCR extensions in the PCR quote and eventlog are not recognized, which then might be used to determine what firmware / OS updates a client has installed, or that it has been compromised)

In this tutorial we'll focus on attestion of servers in an enterprise environment. However, the concepts described here are applicable to other environments, such as IoTs and personal devices, where the attestation database could be hosted on a user's personal devices for use in joining new devices to the user's set of devices, or for joining new IoTs to the user's SOHO network.

Attestation Protocols

Attestation is done by a client computer with a TPM interacting with an attestation service over a network. This requires a network protocol for attestation.

Notation

• Encrypt_<name> == encryption with the named private or secret key (if symmetric, then this primitive is expected to provide authenticated encryption).
• Sign_<name> == digital signature with the named private key.
• MAC_<name> == message authentication code keyed with the named secret key.
• CSn == client-to-server message number n
• SCn == server-to-client message number n
• {stuff, more_stuff} == a sequence of data, a "struct"
• {"key":<value>,...} == JSON text
• TPM2_MakeCredential(<args>) == outputs of calling TPM2_MakeCredential() with args arguments
• TPM2_Certify(<args>) == outputs of calling TPM2_Certify() with args arguments

Proof of Possession of TPM

Proof of possession of a valid TPM is performed by the attestation client sending its TPM's Endorsement Key (EK) certificate (if one is available, else the attestation service must recognize the EK public key) and then exchanging additional messages by which the client can prove its possession of the EK.

Proof of possession of an EK is complicated by the fact that EKs are generally decrypt-only (some TPMs also support signing EKs, but the TCG specifications only require decrypt-only EKs). The protocol has to have the attestation service send a challenge (or key) encrypted to the EKpub and then the attestation client demonstrate that it was able to decrypt that with the EK. However, this is not quite how attestation protocols work! Instead of plain asymmetric encryption the server will use TPM2_MakeCredential(), while the attestation client will use TPM2_ActivateCredential() instead of plain asymmetric decryption.

Trusted State Attestation

Trusted state is attested by sending a quote of Platform Configuration Registers (PCRs) and the eventlog describing the evolution of the system's state from power-up to the current state. The attestation service validates the digests used to extend the various PCRs, and perhaps the sequence in which they appear in the eventlog, typically by checking a list of known-trusted digests (these are, for example, checksums of firmware images).

Typically the attestation protocol will have the client generate a signing-only asymmetric public key pair known as the attestation key (AK) with which to sign the PCR quote and eventlog. Binding of the EKpub and AKpub will happen via TPM2_MakeCredential() / TPM2_ActivateCredential().

Note that the TPM2_Quote() function produces a signed message -- signed with a TPM-resident AK named by the caller (and to which they have access), which would be the AK used in the attestation protocol.

The output of TPM2_Quote() might be the only part of a client's messages to the attestation service that include a signature made with the AK, but integrity protection of everything else can be implied (e.g., the eventlog and PCR values are used to reconstruct the PCR digest signed in the quote). TPM2_Quote() signs more than just a digest of the selected PCRs. TPM2_Quote() signs all of:

• digest of selected PCRs
• caller-provided extra data (e.g., a cookie/nonce/timestamp/...),
• the TPM's firmware version number,
• clock (the TPM's time since startup),
• resetCount (an indirect indicator of reboots),
• restartCount (an indirect indicator of suspend/resume events)
• and safe (a boolean indicating whether the clock might have ever gone backwards).

Binding of Other Keys to EKpub

The semantics of TPM2_MakeCredential() / TPM2_ActivateCredential() make it possible to bind a TPM-resident object to the TPM's EKpub.

TPM2_MakeCredential() encrypts to the EKpub a small secret datum and the name (digest of public part) of the TPM-resident object being bound. The counter-part to this, TPM2_ActivateCredential(), will decrypt that and return the secret to the application IFF (if and only if) the caller has access to the named object.

Typically attestation protocols have the client send its EKpub, EKcert (if it has one), AKpub (the public key of an "attestation key"), and other things (e.g., PCR quote and eventlog signed with the AK), and the server will then send the output of TPM2_MakeCredential() that the client can recover a secret from using TPM2_ActivateCredential().

The implication is that if the client can extract the cleartext payload of TPM2_MakeCredential(), then it must possess a) the EK private key corresponding to the EKpub, b) the AK private key corresponding to the object named by the server.

Proof of possession can be completed immediately by demonstrating knowledge of the secret sent by the server. Proof of possession can also be delayed to an eventual use of that secret, allowing for single round trip attestation.

Binding hosts to TPMs

(TBD. Talk about IDevID or similar certificates binding hosts to their factory-installed TPMs, and how to obtain those from vendors.)

Attestation Protocol Patterns and Actual Protocols (decrypt-only EKs)

Note: all the protocols described below are based on decrypt-only TPM endorsement keys.

Let's start with few observations and security considerations:

• Clients need to know which PCRs to quote. E.g., the Safe Boot project and the IBM sample attestation client and server have the client ask for a list of PCRs and then the client quotes just those.

  But clients could just quote all PCRs. It's more data to send, but probably not a big deal, and it saves a round trip if there's no need to ask what PCRs to send.

• Some replay protection or freshness indication for client requests is needed. A stateful method of doing this is to use a server-generated nonce (as an encrypted state cookie embedding a timestamp). A stateless method is to use a timestamp and reject requests with old timestamps.

• Replay protection of server to client responses is mostly either not needed or implicitly provided by TPM2_MakeCredential() because TPM2_MakeCredential() generates a secret seed that randomizes its outputs even when all the inputs are the same across multiple calls to it.

• Ultimately the protocol must make use of TPM2_MakeCredential() and TPM2_ActivateCredential() in order to authenticate a TPM-running host via its TPM's EKpub.

• Privacy protection of client identifiers may be needed, in which case TLS may be desired.

• Even if a single round trip attestation protocol is adequate, a return routability check may be needed to avoid denial of service attacks. I.e., do not run a single round trip attestation protocol over UDP without first requiring the client to echo a nonce/cookie.

• Statelessness on the server side is highly desirable, as that should permit having multiple servers and each of a client's messages can go to different servers. Conversely, keeping state on the server across multiple round trips can cause resource exhaustion / denial of service attack considerations.

• Statelessness maps well onto HTTP / REST. Indeed, attestation protocol messages could all be idempotent and therefore map well onto HTTP GET requests but for the fact that all the things that may be have to be sent may not fit on a URI local part or URI query parameters, therefore HTTP POST is the better option.

Error Cases Not Shown

Note that error cases are not shown in the protocols described below.

Naturally, in case of error the attestation server will send a suitable error message back to the client.

Databases, Log Sinks, and Dashboarding / Alerting Systems Not Shown

In order to simplify the protocol diagrams below, interactions with databases, log sinks, and alerting systems are not shown.

A typical attestation service will, however, have interactions with those components, some or all of which might even be remote:

• attestation database
• log sinks
• dashboarding / alerting

If an attestation service must be on the critical path for booting an entire datacenter, it may be desirable for the attestation service to be able to run with no remote dependencies, at least for some time. This means, for example, that the attestation database should be locally available and replicated/synchronized only during normal operation. It also means that there should be a local log sink that can be sent to upstream collectors during normal operation.

Single Round Trip Attestation Protocol Patterns

An attestation protocol need not complete proof-of-possession immediately if the successful outcome of the protocol has the client subsequently demonstrate possession to other services/peers. This is a matter of taste and policy. However, one may want to have cryptographically secure "client attested successfully" state on the server without delay, in which case two round trips are the minimum for an attestation protocol.

In the following example the client obtains a certificate (AKcert) for its AKpub, filesystem decryption keys, and possibly other things, and eventually it will use those items in ways that -by virtue of having thus been used- demonstrate that it possesses the EK used in the protocol:

  <client knows a priori what PCRs to quote, possibly all, saving a round trip>

  CS0:  [ID], EKpub, [EKcert], AKpub, PCRs, eventlog, timestamp,
        TPM2_Quote(AK, PCRs, extra_data)=Signed_AK({hash-of-PCRs, misc, extra_data})
  SC0:  {TPM2_MakeCredential(EKpub, AKpub, session_key),
         Encrypt_session_key({AKcert, filesystem_keys, etc.})}

  <extra_data includes timestamp>

  <subsequent client use of AK w/ AKcert, or of credentials made
   available by dint of being able to access filesystems unlocked by
   SC0, demonstrate that the client has attested successfully>

(ID might be, e.g., a hostname.)

(In this diagram we show the use of a TPM simulator on the server side for implementing TPM2_MakeCredential().)

The server will validate that the timestamp is near the current time, the EKcert (if provided, else the EKpub), the signature using the asserted (but not yet bound to the EKpub) AKpub, then it will validate the PCR quote and eventlog, and, if everything checks out, will issue a certificate for the AKpub and return various secrets that the client may need.

The client obtains those items IFF (if and only if) the AK is resident in the same TPM as the EK, courtesy of TPM2_ActivateCredential()'s semantics.

NOTE well that in single round trip attestation protocols using only decrypt-only EKs it is essential that the AKcert not be logged in any public place since otherwise an attacker can make and send CS0 using a non-TPM-resident AK and any TPM's EKpub/EKcert known to the attacker, and then it may recover the AK certificate from the log in spite of being unable to recover the AK certificate from SC1!

Alternatively, a single round trip attestation protocol can be implemented as an optimization to a two round trip protocol when the AK is persisted both, in the client TPM and in the attestation service's database:

  <having previously successfully enrolled AKpub and bound it to EKpub...>

  CS0:  timestamp, AKpub, PCRs, eventlog,
        TPM2_Quote(AK, PCRs, extra_data)=Signed_AK({hash-of-PCRs, misc, extra_data})
  SC0:  {TPM2_MakeCredential(EKpub, AKpub, session_key),
         Encrypt_session_key({AKcert, filesystem_keys, etc.})}

Three-Message Attestation Protocol Patterns

A single round trip protocol using encrypt-only EKpub will not demonstrate proof of possession immediately, but later on when the certified AK is used elsewhere. A proof-of-possession (PoP) may be desirable anyways for monitoring and alerting purposes.

  CS0:  [ID], EKpub, [EKcert], AKpub, PCRs, eventlog, timestamp,
        TPM2_Quote(AK, PCRs, extra_data)=Signed_AK({hash-of-PCRs, misc, extra_data})
  SC0:  {TPM2_MakeCredential(EKpub, AKpub, session_key),
         Encrypt_session_key({AKcert, filesystem_keys, etc.})}
  CS1:  AKcert, Signed_AK(AKcert)

(In this diagram we show the use of a TPM simulator on the server side for implementing TPM2_MakeCredential().)

NOTE well that in this protocol, like single round trip attestation protocols using only decrypt-only EKs, it is essential that the AKcert not be logged in any public place since otherwise an attacker can make and send CS0 using a non-TPM-resident AK and any TPM's EKpub/EKcert known to the attacker, and then it may recover the AK certificate from the log in spite of being unable to recover the AK certificate from SC1!

If such a protocol is instantiated over HTTP or TCP, it will really be more like a two round trip protocol:

  CS0:  [ID], EKpub, [EKcert], AKpub, PCRs, eventlog, timestamp,
        TPM2_Quote(AK, PCRs, extra_data)=Signed_AK({hash-of-PCRs, misc, extra_data})
  SC0:  {TPM2_MakeCredential(EKpub, AKpub, session_key),
         Encrypt_session_key({AKcert, filesystem_keys, etc.})}
  CS1:  AKcert, Signed_AK(AKcert)
  SC1:  <empty>

Two Round Trip Stateless Attestation Protocol Patterns

We can add a round trip to the protocol in the previous section to make the client prove possession of the EK and binding of the AK to the EK before it can get the items it needs. This avoids the security consideration of having to not log the AKcert.

Below is a sketch of a stateless, two round trip attestation protocol.

Actual protocols tend to use a secret challenge that the client echoes back to the server rather than a secret key possesion of which is proven with symmetriclly-keyed cryptographic algorithms.

  CS0:  [ID], EKpub, [EKcert], AKpub, PCRs, eventlog, timestamp,
        TPM2_Quote(AK, PCRs, extra_data)=Signed_AK({hash-of-PCRs, misc, extra_data})
  SC0:  {TPM2_MakeCredential(EKpub, AKpub, session_key), ticket}
  CS1:  {ticket, MAC_session_key(CS0), CS0}
  SC1:  Encrypt_session_key({AKcert, filesystem_keys, etc.})

  <extra_data includes timestamp>

where session_key is an ephemeral secret symmetric authenticated encryption key, and ticket is an authenticated encrypted state cookie:

  ticket = {vno, Encrypt_server_secret_key({session_key, timestamp,
                                            MAC_session_key(CS0)})}

where server_secret_key is a key known only to the attestation service and vno identifies that key (in order to support key rotation without having to try authenticated decryption twice near key rotation events).

[Note: ticket here is not in the sense used by TPM specifications, but in the sense of "TLS session resumption ticket" or "Kerberos ticket", and, really, it's just an encrypted state cookie so that the server can be stateless.]

The attestation server could validate that the timestamp is recent upon receipt of CS0. But the attestation server can delay validation of EKcert, signatures, and PCR quote and eventlog until receipt of CS1. In order to produce SC0 the server need only digest the AKpub to produce the name input of TPM2_MakeCredential(). Upon receipt of CS1 (which repeats CS0), the server can decrypt the ticket, validate the MAC of CS0, validate CS0, and produce SC1 if everything checks out.

In this protocol the client must successfully call TPM2_ActivateCredential() to obtain the session_key that it then proves possession of in CS1, and only then does the server send the AKcert and/or various secret values to the client, this time saving the cost of asymmetric encryption by using the session_key to key a symmetric authenticated cipher.

(The server_secret_key, ticket, session_key, and proof of possession used in CS1 could even conform to Kerberos or encrypted JWT and be used for authentication, possibly with an off-the-shelf HTTP stack.)

An HTTP API binding for this protocol could look like:

  POST /get-attestation-ticket
      Body: CS0
      Response: SC0

  POST /attest
      Body: CS1
      Response: SC1

Here the attestation happens in the first round trip, but the proof of possession is completed in the second, and the delivery of secrets and AKcert also happens in the second round trip.

Actual Protocols: ibmacs

The IBM TPM Attestation Client Server (ibmacs) open source project has sample code for a "TCG attestation application".

It implements a stateful (state is kept in a database) attestation and enrollment protocol over TCP sockets that consists of JSON texts of the following form, sent prefixed with a 32-bit message length in host byte order:

  CS0: {"command":"nonce","hostname":"somehostname",
        "userid":"someusername","boottime":"2021-04-29 16:37:06"}
  SC0: {"response":"nonce","nonce":"<hex>", "pcrselect":"<hex>", ...}

  <nonce is used in production of signed PCR quote>

  CS1: {"command":"quote","hostname":"somehostname",
        "quoted":"<hex>","signature":"<hex>",
        "event1":"<hex>","imaevent0":"<hex>"}
  SC1: {"response":"quote"}

  CS2: {"command":"enrollrequest","hostname":"somehost",
        "tpmvendor":"...","ekcert":"<PEM>","akpub":"<hex(DER)>"}
  SC2: {"response":"enrollrequest",
        "credentialblob":"<hex of credentialBlob output of TPM2_MakeCredential()>",
        "secret":"<hex of secret output of TPM2_MakeCredential()>"}

  CS3: {"command":"enrollcert","hostname":"somecert","challenge":"<hex>"}
  SC3: {"response":"enrollcert","akcert":"<hex>"}

The server keeps state across round trips.

Note that this protocol has up to four (4) round trips. Because the ibmacs server keeps state in a database, it should be possible to elide some of these round trips in attestations subsequent to enrollment.

The messages of the second and third round trips could be combined since there should be no need to wait for PCR quote validation before sending the EKcert and AKpub. The messages of the first round trip too could be combined with the messages of the second and third round trip by using a timestamp as a nonce -- with those changes this protocol would get down to two round trips.

Actual Protocols: safeboot.dev

  CS0:  <empty>
  SC0:  nonce, PCR_list
  CS1:  [ID], EKpub, [EKcert], AKpub, PCRs, eventlog, nonce,
        TPM2_Quote(AK, PCRs, extra_data)=Signed_AK({hash-of-PCRs, misc, extra_data})
  SC1:  {TPM2_MakeCredential(EKpub, AKpub, session_key),
         Encrypt_session_key({filesystem_keys})}

Nonce validation is currently not well-developed in Safeboot. If a timestamp is used instead of a nonce, and if the client assumes all PCRs are desired, then this becomes a one round trip protocol.

An AKcert will be added to the Safeboot protocol soon.

Actual Protocols: ...

(TBD)

Attestation Protocol Patterns and Actual Protocols (signing-only EKs)

Some TPMs come provisioned with signing-only endorsement keys in addition to decrypt-only EKs. For example, vTPMs in Google cloud provides both, decrypt-only and signing-only EKs.

Signing-only EKs can be used for attestation as well.

[Ideally signing-only EKs can be restricted to force the use of TPM2_Certify()? Restricted signing keys can only sign payloads that start with a magic value, whereas unrestricted signing keys can sign any payload.]

Signing-only EKs make single round trip attestation protocols possible that also provide immediate attestation status because signing provides proof of possession non-interactively, whereas asymmetric encryption requires interaction to prove possession:

  CS0:  timestamp, [ID], EKpub, [EKcert], AKpub, PCRs, eventlog,
        TPM2_Certify(EKpub, AKpub), TPM2_Quote()
  SC0:  AKcert

If secrets need to be sent back, then a decrypt-only EK also neds to be used:

  CS0:  timestamp, [ID], EKpub_signing, EKpub_encrypt,
        [EKcert_signing], [EKcert_encrypt], AKpub, PCRs, eventlog,
        TPM2_Certify(EKpub, AKpub), TPM2_Quote()
  SC0:  {TPM2_MakeCredential(EKpub_encrypt, AKpub, session_key),
         Encrypt_session_key({AKcert, filesystem_keys, etc.})}

Long-Term State Kept by Attestation Services

Attestation servers need to keep some long-term state:

• binding of EKpub and ID
• PCR validation profile(s) for each identified client
• resetCount (for reboot detection)

Log-like attestation state:

• client attestation status (last time successfully attested, last time unsuccessfully attested)

The PCR validation profile for a client consists of a set of required and/or acceptable digests that must appear in each PCR's extension log. These required and/or acceptable digests may be digests of firmware images, boot loaders, boot loader configurations (e.g., menu.lst, for Grub), operating system kernels, initrd images, filesystem root hashes (think ZFS), etc.

Some of these are obtained by administrators on a trust-on-first-use (TOFU) basis.

Things to log:

• client attestation attempts and outcomes
• AK certificates issued (WARNING: see note about single round trip attestation protocols above -- do not log AKcerts in public places when using single round trip attestation protocols!)

Long-Term State Created or Updated by Attestation Services

• An attestation service might support creation of host<->EKpub bindings on a first-come-first-served basis. In this mode the attestation server might validate an EKcert and that the desired hostname has not been bound to an EK, then create the binding.

• An attestation service might support deletion of host PCR validation profiles that represent past states upon validation of PCR quotes using newer profiles. This could be used to permit firmware and/or operating system upgrades and then disallow downgrades after evidence of successful upgrade.

• An attestation service might keep track of client reboots so as to:

  • revoke old AKcerts when the client reboots (but note that this is really not necessary if we trust the client's TPM, since then the previous AKs will never be usable again)
  • alert if the reboot count ever goes backwards

Schema for Attestation Server Database

A schema for the attestation server's database entries might look like:

{
  "EKpub": "<EKpub>",
  "hostname": "<hostname>",
  "EKcert": "<EKcert in PEM, if available>",
  "previous_firmware_profile": "FWProfile0",
  "current_firmware_profiles": ["FWProfile1", "FWProfile2", "..."],
  "previous_operating_system_profiles": "OSProfile0",
  "current_operating_system_profiles": ["OSProfile1", "OSProfile2", "..."],
  "previous_PCRs": "<...>",
  "proposed_PCRs": "<...>",
  "ak_cert_template": "<AKCertTemplate>",
  "secrets": "<secrets>",
  "resetCount": "<resetCount value from last quote>"
}

The attestation server's database should have two lookup keys:

• EKpub
• hostname

The attestation server's database's entry for any client should provide, de minimis:

• a way to validate the root of trust measurements in the client's quoted PCRs, for which two methods are possible:
  • save the PCRs quoted last as the ones expected next time
  • or, name profiles for validating firmware RTM PCRs and profiles for validating operating system RTM PCRs

A profile for validating PCRs should contain a set of expected extension values for each of a set of PCRs. The attestation server can then check that the eventlog submitted by the client lists exactly those extension values and no others. PCR extension order in the eventlog probably doesn't matter here. If multiple profiles are named, then one of those must match -- this allows for upgrades and downgrades.

{
  "profile_name":"SomeProfile",
  "values":[
    {
      "PCR":0,
      "values":["aaaaaaa","bbbbbb","..."]
    },
    {
      "PCR":1,
      "values":["ccccccc","dddddd","..."]
    }
  ]
}

Using the PCR values from the previous attestation makes upgrades tricky, probably requiring an authenticated and authorized administrator to bless new PCR values after an upgrade. A client that presents a PCR quote that does not match the previous one would cause the proposed_PCRs field to be updated but otherwise could not continue, then an administrator would confirm that the client just did a firmware/OS upgrade and if so replace the previous_PCRs with the proposed_PCRs, then the client could attempt attestation again.

Dealing with Secrets

An attestation server might want to return storage/filesystem decryption key-encryption-keys to a client. But one might not want to store those keys in the clear on the attestation server. As well, one might want a break-glass way to recover those secrets.

For break-glass recovery, the simplest thing to do is to store Encrypt_backupKey({EKpub, hostname, secrets}), where backupKey is an asymmetric key whose private key is stored offline (e.g., in a safe, or in an offline HSM). To break the glass and recover the key, just bring the ciphertext to the offline system where the private backup key is kept, decrypt it, and then use the secrets manually to recover the affected system.

Here are some ideas for how to make an attestation client depend on the attestation server giving it keys needed to continue booting after successful attestation:

• Store TPM2_MakeCredential(EKpub, someObjectName, key0), Encrypt_key0(secrets).

  In this mode the server sends the client the stored data, then client gets to recreate someObject (possibly by loading a saved object or by re-creating it on the same non-NULL hierarchy from the same primary seed using the same template and extra entropy) on its TPM so that the corresponding call to TPM2_ActivateCredential() can succeed, then the client recovers key0 and decrypts the encrypted secrets. Here someObject can be trivial and need only exist to make the {Make,Activate}Credential machinery work.

  TPM replacement and/or migration of a host from one physical system to another can be implemented by learning the new system's TPM's EKpub and using the offline backupKey to compute TPM2_MakeCredential(EKpub_new, someObjectName, key0) and update the host's entry.

• Alternatively generate a non-restricted decryption private key using a set template and extra entropy, on the same non-NULL hierarchy (i.e., from the same seed), enroll the public key to this private key in an attestation protocol, and have the attestation server store secrets encrypted to that public key.

  (The EK cannot be used this way because it is restricted.)

• Store a secret value that will be extended into an application PCR that is used as a policy PCR for unsealing a persistent object stored on the client's TPM.

  In this mode the server sends the client the secret PCR extension value, and the client uses it to extend a PCR such that it can then unseal the real storage / filesystem decryption keys.

  Using a PCR and a policy on the key object allows for a clever break-glass secret recovery mechanism by using a compound extended authorization (EA) policy that allows either unsealing based on a PCR, or maybe based on an password-based HMAC (with machine passwords stored in a safe).

• A hybrid of the previous options, where the server stores a secret PCR extension value wrapped with TPM2_MakeCredential().

Other ideas?

References

• TCG TPM Library part 1: Architecture, sections 23 and 24
• https://sourceforge.net/projects/ibmtpm20acs/
• https://safeboot.dev/
• https://github.com/osresearch/safeboot/