I finally figured it out myself. It should come as no surprise, but the answer is nauseatingly complex and indirect.
The missing pieces to the puzzle were in RFC 5652. I didn't really understand the TimeStampResp structure until I read (well, skimmed through) that document.
Let me describe in brief the TimeStampReq and TimeStampResp structures. The interesting fields of the request are:
- a "message imprint", which is the hash of the data to be timestamped
- the OID of the hash algorithm used to create the message imprint
- an optional "nonce", which is a client-chosen identifier used to verify that the response is generated specifically for this request. This is effectively just a salt, used to avoid replay attacks and to detect errors.
The meat of the response is a CMS SignedData structure. Among the fields in this structure are:
- the certificate(s) used to sign the response
- an EncapsulatedContentInfo member containing a TSTInfo structure. This structure, importantly, contains:
- the message imprint that was sent in the request
- the nonce that was sent in the request
- the time certified by the TSA
- a set of SignerInfo structures, with typically just one structure in the set. For each SignerInfo, the interesting fields within the structure are:
- a sequence of "signed attributes". The DER-encoded BLOB of this sequence is what is actually signed. Among these attributes are:
- the time certified by the TSA (again)
- a hash of the DER-encoded BLOB of the TSTInfo structure
- an issuer and serial number or subject key identifier that identifies the signer's certificate from the set of certificates found in the SignedData structure
- the signature itself
- a sequence of "signed attributes". The DER-encoded BLOB of this sequence is what is actually signed. Among these attributes are:
The basic process of validating the timestamp is as follows:
- Read the data that was timestamped, and recompute the message imprint using the same hashing algorithm used in the timestamp request.
- Read the nonce used in the timestamp request, which must be stored along with the timestamp for this purpose.
- Read and parse the TimeStampResp structure.
- Verify that the TSTInfo structure contains the correct message imprint and nonce.
- From the TimeStampResp, read the certificate(s).
- For each SignerInfo:
- Find the certificate for that signer (there should be exactly one).
- Verify the certificate.
- Using that certificate, verify the signer's signature.
- Verify that the signed attributes contain the correct hash of the TSTInfo structure
If everything is okay, then we know that all signed attributes are valid, since they're signed, and since those attributes contain a hash of the TSTInfo structure, then we know that's okay, too. We have therefore validated that the timestamped data is unchanged since the time given by the TSA.
Because the signed data is a DER-encoded BLOB (which contains a hash of the different DER-encoded BLOB containing the information the verifier actually cares about), there's no getting around having some sort of library on the client (verifier) that understands X.690 encoding and ASN.1 types. Therefore, I conceded to including Bouncy Castle in the client as well as in the build process, since there's no way I have time to implement those standards myself.
My code to add and verify timestamps is similar to the following:
Timestamp generation at build time:
// a lot of fully-qualified type names here to make sure it's clear what I'm using
static void WriteTimestampToBuild(){
var dataToTimestamp = ... // see OP
var hashToTimestamp = ... // see OP
var nonce = ... // see OP
var tsq = GetTimestampRequest(hashToTimestamp, nonce);
var tsr = GetTimestampResponse(tsq, "http://some.rfc3161-compliant.server");
ValidateTimestamp(tsq, tsr);
WriteToBuild("tsq-hashalg", Encoding.UTF8.GetBytes("SHA1"));
WriteToBuild("nonce", nonce.ToByteArray());
WriteToBuild("timestamp", tsr.GetEncoded());
}
static Org.BouncyCastle.Tsp.TimeStampRequest GetTimestampRequest(byte[] hash, Org.BouncyCastle.Math.BigInteger nonce){
var reqgen = new TimeStampRequestGenerator();
reqgen.SetCertReq(true);
return reqgen.Generate(TspAlgorithms.Sha1/*assumption*/, hash, nonce);
}
static void GetTimestampResponse(Org.BouncyCastle.Tsp.TimeStampRequest tsq, string url){
// similar to OP
}
static void ValidateTimestamp(Org.BouncyCastle.Tsp.TimeStampRequest tsq, Org.BouncyCastle.Tsp.TimeStampResponse tsr){
// same as client code, see below
}
static void WriteToBuild(string key, byte[] value){
// not shown
}
Timestamp verification at run time (client site):
/* Just like in the OP, I've used fully-qualified names here to avoid confusion.
* In my real code, I'm not doing that, for readability's sake.
*/
static DateTime GetTimestamp(){
var timestampedData = ReadFromBuild("timestamped-data");
var hashAlg = Encoding.UTF8.GetString(ReadFromBuild("tsq-hashalg"));
var timestampedHash = System.Security.Cryptography.HashAlgorithm.Create(hashAlg).ComputeHash(timestampedData);
var nonce = new Org.BouncyCastle.Math.BigInteger(ReadFromBuild("nonce"));
var tsq = new Org.BouncyCastle.Tsp.TimeStampRequestGenerator().Generate(System.Security.Cryptography.CryptoConfig.MapNameToOID(hashAlg), timestampedHash, nonce);
var tsr = new Org.BouncyCastle.Tsp.TimeStampResponse(ReadFromBuild("timestamp"));
ValidateTimestamp(tsq, tsr);
// if we got here, the timestamp is okay, so we can trust the time it alleges
return tsr.TimeStampToken.TimeStampInfo.GenTime;
}
static void ValidateTimestamp(Org.BouncyCastle.Tsp.TimeStampRequest tsq, Org.BouncyCastle.Tsp.TimeStampResponse tsr){
/* This compares the nonce and message imprint and whatnot in the TSTInfo.
* It throws an exception if they don't match. This doesn't validate the
* certs or signatures, though. We still have to do that in order to trust
* this data.
*/
tsr.Validate(tsq);
var tst = tsr.TimeStampToken;
var timestamp = tst.TimeStampInfo.GenTime;
var signers = tst.ToCmsSignedData().GetSignerInfos().GetSigners().Cast<Org.BouncyCastle.Cms.SignerInformation>();
var certs = tst.GetCertificates("Collection");
foreach(var signer in signers){
var signerCerts = certs.GetMatches(signer.SignerID).Cast<Org.BouncyCastle.X509.X509Certificate>().ToList();
if(signerCerts.Count != 1)
throw new Exception("Expected exactly one certificate for each signer in the timestamp");
if(!signerCerts[0].IsValid(timestamp)){
/* IsValid only checks whether the given time is within the certificate's
* validity period. It doesn't verify that it's a valid certificate or
* that it hasn't been revoked. It would probably be better to do that
* kind of thing, just like I'm doing for the signing certificate itself.
* What's more, I'm not sure it's a good idea to trust the timestamp given
* by the TSA to verify the validity of the TSA's certificate. If the
* TSA's certificate is compromised, then an unauthorized third party could
* generate a TimeStampResp with any timestamp they wanted. But this is a
* chicken-and-egg scenario that my brain is now too tired to keep thinking
* about.
*/
throw new Exception("The timestamp authority's certificate is expired or not yet valid.");
}
if(!signer.Verify(signerCerts[0])){ // might throw an exception, might not ... depends on what's wrong
/* I'm pretty sure that signer.Verify verifies the signature and that the
* signed attributes contains a hash of the TSTInfo. It also does some
* stuff that I didn't identify in my list above.
* Some verification errors cause it to throw an exception, some just
* cause it to return false. If it throws an exception, that's great,
* because that's what I'm counting on. If it returns false, let's
* throw an exception of our own.
*/
throw new Exception("Invalid signature");
}
}
}
static byte[] ReadFromBuild(string key){
// not shown
}