It makes sense to support multiple encryption backends. So, there should be a way to tell what backend is responsible for a given filename in an encrypted remote. (And since special remotes can also store files unencrypted, differentiate from those as well.)
The rest of this page will describe a single encryption backend using GPG. Probably only one will be needed, but who knows? Maybe that backend will turn out badly designed, or some other encryptor needed. Designing with more than one encryption backend in mind helps future-proofing.
The basis of this scheme was originally developed by Lars Wirzenius et al for Obnam.
Data is encrypted by GnuPG, using a symmetric cipher. The cipher is
generated by GnuPG when the special remote is created. By default the
best entropy pool is used, hence the generation may take a while; One
to speed up things, but at the expense of using random numbers of a
lower quality. The generated cipher is then checked into your git
repository, encrypted using one or more OpenPGP public keys. This scheme
allows new OpenPGP private keys to be given access to content that has
already been stored in the remote.
Different encrypted remotes need to be able to each use different ciphers. Allowing multiple ciphers to be used within a single remote would add a lot of complexity, so is not planned to be supported. Instead, if you want a new cipher, create a new S3 bucket, or whatever. There does not seem to be much benefit to using the same cipher for two different encrypted remotes.
So, the encrypted cipher could just be stored with the rest of a remote's
remotes.log (see internals). When
annex intiremote makes a remote, it can generate a random symmetric
cipher, and encrypt it with the specified gpg key. To allow another gpg
public key access, update the encrypted cipher to be encrypted to both gpg
If the names of files are encrypted or securely hashed, or whatever is chosen, this makes it harder for git-annex (let alone untrusted third parties!) to get a list of the files that are stored on a given enrypted remote. But, does git-annex really ever need to do such an enumeration?
git annex unused --from remote can now check for
unused data that is stored on a remote, and it does so based only on
location log data for the remote. This assumes that the location log is
git annex fsck --from remote? Such a command should be able to,
for each file in the repository, contact the encrypted remote to check
if it has the file. This can be done without enumeration, although it will
mean running gpg once per file fscked, to get the encrypted filename.
So, the files stored in the remote should be encrypted. But, it needs to be a repeatable encryption, so they cannot just be gpg encrypted, that would yeild a new name each time. Instead, HMAC is used. Any hash could be used with HMAC. SHA-1 is the default, but other hashes can be chosen for new remotes.
It was suggested that it might not be wise to use the same cipher for both gpg and HMAC. Being paranoid, it's best not to tie the security of one to the security of the other. So, the encrypted cipher described above is actually split in two; half is used for HMAC, and half for gpg.
Does the HMAC cipher need to be gpg encrypted? Imagine if it were stored in plainext in the git repository. Anyone who can access the git repository already knows the actual filenames, and typically also the content hashes of annexed content. Having access to the HMAC cipher could perhaps be said to only let them verify that data they already know.
While this seems a pretty persuasive argument, I'm not 100% convinced, and anyway, most times that the HMAC cipher is needed, the gpg cipher is also needed. Keeping the HMAC cipher encrypted does slow down two things: dropping content from encrypted remotes, and checking if encrypted remotes really have content. If it's later determined to be safe to not encrypt the HMAC cipher, the current design allows changing that, even for existing remotes.
The symmetric cipher can be used to encrypt other content than the content sent to the remote. In particular, it may make sense to encrypt whatever access keys are used by the special remote with the cipher, and store that in remotes.log. This way anyone whose gpg key has been given access to the cipher can get access to whatever other credentials are needed to use the special remote.
A risk of this scheme is that, once the symmetric cipher has been
obtained, it allows full access to all the encrypted content. Indeed
anyone owning a key that used to be granted access could already have
decrypted the cipher and stored a copy. While it is in possible to
remove a key with
keyid-=, it is designed for a
completely different purpose and does not actually revoke
If git-annex stores the decrypted symmetric cipher in memory, then there is a risk that it could be intercepted from there by an attacker. Gpg ameliorates these type of risks by using locked memory. For git-annex, note that an attacker with local machine access can tell at least all the filenames and metadata of files stored in the encrypted remote anyway, and can access whatever content is stored locally.
This design does not support obfuscating the size of files by chunking them, as that would have added a lot of complexity, for dubious benefits. If the untrusted party running the encrypted remote wants to know file sizes, they could correlate chunks that are accessed together. Encrypting data changes the original file size enough to avoid it being used as a direct fingerprint at least.