Using EKhash = H(EKpub), when the EKpub is derived from an EKcert, is problematic

[nicowilliams]

In HCP we use Safeboot w/ an EKcert, then we derive an EKpub from the EKcert, then we compute a hash of the EKpub and name the enrollment after that hash. At attestation time the client sends its EKpub and the server looks up the enrollment by EKhash = H(EKpub).

The way the EKcert->EKpub conversion is done is simple:

• validate the EKcert
• extract the subject public key from the EKcert
• load that public key into an swTPM as a public key object with all the attributes expected of an EK (the standard authPolicy, attributes, etc.)
• extract the TPM2B_PUBLIC for that loaded object from the swTPM

But there are at least two ways in which the TPM2B_PUBLIC of an EK as produced by the TPM where it resides, and as produced by another TPM as used above can differ:

• if the key is an RSA key, the exponent should be 2^16 + 1, which is the default exponent, and which should be written as the all-zeros value that indicates the default The swTPM writes all-zeros. A TPM could write 2^16 + 1, and at least one brand does.

• if the key is a restricted, decrypt-only RSA key, then the scheme is required to be OEAP and that is allowed to be expressed as the default scheme (NULL) We've not observed any TPM writing in the OEAP scheme instead of the NULL scheme, but it could happen.

The upshot is that either we must

• enroll both, an EKcert and an EKpub (checking that the spk matches, naturally),
• send the EKcert during attestation, or
• normalize the the EKpub sent by the client at attestation time.

(Readers should note that enrollment also supports enrolling with an EKpub directly, and this is for the case where virtual TPMs are used where there is a mechanism other than EKcerts for validating the legitimacy of a TPM.)

[nicowilliams]

I think the normalization approach will probably be the best.

[osresearch]

Normalizing the EKpub on the server might make more sense? The server is the one choosing to use the hash, not the client, and a different server might choose an equality test or something else like checking the modulus, which means that the client doesn't know what fields are important.

This normalization is also easier to do in the Python server than the shell scripts of the client, right?

[nicowilliams]

This normalization is also easier to do in the Python server than the shell scripts of the client, right?

I'm thinking it's just tpm2_loadexternal then tpm2_readpublic to normalize.

The swtpm does remember whether the exponent was given as 0x00010001 or 0x00000000, so we'll have to resort to ugly games.

A dd command works. However, I fear this is too ugly, and that we might need to do something else.

Options:

• hash just the raw public key from the EKpub, not its full TPM2B_PUBLIC
• require the EKpub in addition to the EKcert at enrollment time (this might not be easy to pull off for organizations running enrollment)