Incremental backup with strong cryptographic confidentiality baked into the data model. In a small package, with no dependencies.

This project is still experimental! Things may break or change. See below on status.

• Designed around public key cryptography such that decryption key can be kept offline, air-gapped.

• Backup to local or remote storage with arbitrary transport.

• Incremental update built on inode identity and hashed block contents, compatible with moving and reorganizing entire trees.

• Data deduplication.

• Low local storage requirements for change tracking -- roughly 56-120 bytes per file plus 0.1-5% of total data size.

• Live-streamable to storage. Compatible with append-only media. No local storage required for staging a backup that will be stored remotely.

• Optional support for blinded garbage-collection of blobs on the storage host side.

• Written entirely in C with no library dependencies. Requires no installation.

• Built on modern cryptographic primitives: Curve25519 ECDH, ChaCha20, and SHA-3.

Bakelite is presently experimental and is a work in progress. The above-described features are all present, but have not been subjected to third party review or extensive testing. Moreover, many advanced features normally expected in backup software, like to controls over inclusion/exclusion of files, are not yet available. The codebase is also in transition from a rapidly developed proof of concept to something more mature and well-factored.

Data formats may be subject to change. If attempting to use Bakelite as part of a real backup workflow, you should keep note of the particular version used in case it's needed for restore. Note that the actual backup format is more mature and stable than the configuration and local index format, so the more likely mode of breakage when upgrading is needing to start a new full (non-incremental) backup, not inability to read old backups.

Backups are inherently attack surface on the privacy/confidentiality of one's data. For decades I've looked for a backup system with the right cryptographic properties to minimize this risk, and turned up nothing, leaving me reliant on redundant copies of important things (the "Linus backup strategy") rather than universal system-wide backups. After some moderately serious data loss, I decided it was time to finally write what I had in mind.

Among the near-solutions out there, some required large library or language runtime dependencies, some only worked with a particular vendor's cloud storage service, and some had poor change tracking that failed to account for whole trees being moved or renamed. But most importantly, none with incremental capability addressed the catastrophic loss of secrecy of all past and current data in the event that the encryption key was exposed.

A backup image is a Merkle tree of nodes representing directories, files, and file content blocks, with each node identified by a SHA-3 hash of its encrypted contents, and the root of the tree referenced by a signed summary record. For readers familiar with the git data model, this is very much like a git tree (not commit) but with the objects encrypted. Multiple trees can share common subtrees. This is how incremental backups are represented, and is analogous to how git commits share subtrees. Backup snapshots are not like git commits however; they do not reference each other or have parent/child relationships. This allows arbitrary retention policies to be implemented without breaking any Merkle tree reference chains.

Since there is no way for the system being backed up to "read back" from the backups when it doesn't hold the private decryption key, a "local index" is kept to track how objects in storage correspond to the live filesystem contents. It is a key/value dictionary mapping (device,inode) pairs and hashes of unencrypted file content blocks to the corresponding encrypted object hashes. The local index does not need to be stored with the backup, and should not be. A party who has read access to the local index can probe whether known data was present on the filesystem at the time of last backup, which inode(s) (thereby which files, if they exist in listable directories) contained that data, and which inode(s) share common contents. (Note that these are exactly the capabilities needed for deduplicating of data within and between snapshots.)

• If neither private nor public key is exposed (perspective of backup storage provider), breaking confidentiality of backup depends on breaking ChaCha20.

• If the public key is exposed (for example, via breach on the system being backed up), breaking confidentiality of backup depends on breaking ChaCha20 or solving the computational discrete logarithm problem on Curve25519.

• Breaking integrity of backup depends on breaking second-preimage resistance of SHA-3 or breaking the signing algorithm used (signature forgery). The latter admits only complete tree replacement, not selective modification.

1. Key generation. This step does not need to be done on the system that will be backed up, and should be done on a system you absolutely trust -- both to have a working cryptographic entropy source, and not to expose data. Choose a place to store the secret key, such as an encrypted removable device, and run:

    bakelite genkey backup.sec
    bakelite pubkey backup.sec > backup.pub
   

   Then copy backup.pub to the system(s) you want to back up using this key.

2. Initialization. On the system to be backed up, create an empty directory and run:

    bakelite init /path/to/backup.pub /path/needing/backup
   

   This will create a skeleton configuration in the current working directory. All further steps should be performed from this directory.

3. Configure storage. Edit the store_cmd script produced by bakelite init to something that will accept data in tar format and write it to the desired storage, reporting success or failure via exit status. For example, for local storage to mounted media:

    tar -C /media/backup -kxf -
   

   or appending to a tape drive:

    cat >> /dev/nst0
   

   or to a remote host via ssh:

    ssh backup@remotehost
   

   In the latter (ssh) case, the remote authorized_keys file should force a command that stores the tar stream appropriately and disallows overwrite of existing data. For example:

    command="tar -C /media/backup -kxf -" ssh-ed25519 AAAA...
   

4. Configure devices. Normally, Bakelite will not traverse mount points to other devices; this avoids accidentally including transient mounts of external media or remote shares into a backup they don't belong in. If you want to include additional mounts, create a symlink to the root of each in the directory named "devices". The symlink name will serve as a "label" for the device used in the local indexing, so that changes to device numbering across reboots do not break the index. For example:

    ln -s /home devices/home
    ln -s /var devices/var
   

5. Configure signatures. Create an executable sign_cmd file that accepts data to sign on stdin and produces a signature file on stdout. For example, to use signify:

    signify -S -s signing.sec -x - -m -
   

6. Additional configuration. Edit the config file to change any other preferences as desired. It's recommended to at least set a label for the backup so that the signed summary files will be associated with a particular role/identity, unless separate signing keys will be used for each tree being backed up.

   To exclude files matching certain patterns from backups, create a file named exclude containing one pattern per line. Patterns are a superset of standard glob pattern functionality, intended to match .gitignore conventions, except that inversion using leading ! is not supported. In particular, ** can be used to match zero or more path components, final / forces only directories to match, and patterns with no / (except possibly a final one) can match in any directory (they have an implicit **/ prefix).

   If the directory containing backup configuration is included in the backup, it is recommended to exclude index* from this directory, since the index will be out-of-date at the time of backup and index.pending will be incomplete. Instead of backing it up, the index file can be recreated at restore time if desired.

7. Run the first backup.

    bakelite backup -v
   

   The -v (verbose) flag is helpful to see what's happening, especially for new or changed configurations. However, it does expose information about filesystem contents/changes. Setups aiming to maximize privacy should not use it in an automated setting with logging.

   When the job is finished, a text file named according to the label and UTC backup timestamp, in the form label-yyyy-mm-ddThhmmss.nnnnnnnnnZ.txt, should be present on the backage storage medium, along with a number of files with hex string names in the objects directory. A .sig file will be present too if signing was configured.

8. Setup a cron job to perform further backups on the desired schedule. For example:

    0 2 * * * bakelite -C /path/to/configuration/dir backup
   

It's recommended to test that you are able to restore backups. On a system with the secret key available, run the restore command, as in:

bakelite restore -v -k backup.sec -d /dest/path summary_file.txt

If the secret key is protected by a passphrase you will need to enter it. (Note: passphrase-protected key files are not yet implemented.)

By default, objects are searched for in objects/ relative to the location of the summary file (the same as the default tree layout in the tar stream emitted by the backup command for storage).

If you wish to continue incremental use of the backup after restore, you will need to rebuild the local index as part of the restore operation, using the -i index option.

During testing, original and restored trees should be compared, either directly with a tool like diff or by recursively printing hashes and metadata with find, ls -lR, or similar and diffing the output, to satisfy oneself that the backup was faithful and faithfully restored.

Note that non-POSIX metadata such as extended attributes is not yet stored in the backup inode records or restored. Support for this functionality may be added in the future.

It's entirely possible to treat the backup storage as append-only, cycling media and performing a new full backup periodically or when the media fills up. However, it's also possible to use a running incremental backup indefinitely without filling up storage, by scripting a retention policy to delete old summary files and and prune (garbage collect) data objects that the remaining snapshots do not reference. This can be done without access to the backup contents (i.e. without the private key) via bloom filters attached to each summary.

From the directory containing the summary records and objects directory, run:

bakelite prune *.txt

This will output a list of relative object file pathnames which are not referenced by any of *.txt, which can be fed into xargs to actually delete them.

Bakelite includes simple retention policy for backup summaries via the cull subcommand:

bakelite cull -d7 -w4 -m4 -y1 *.txt

will scan the timestamps in the provided files (*.txt) and output a list of which can be deleted while still keeping the latest from each of the last 7 days, 4 weeks, 4 months, and 1 year. In addition, all in a given range of seconds (default 86400, selected by -r) before the latest will be kept. As with prune, the output of bakelite cull can be piped into xargs for actual deletion (or moving files to stage them for deletion).

Eventually, cull will include label matching and signature checking to ensure that untrusted or erroneously included files from another backup are not used in computing the results, but this functionality is not yet implemented.

By scheduling the appropriate cull and prune commands on a backup storage host, continuous incremental backups can be kept in bounded storage space.