seccs 0.0.5

Secure Content Store

What it is

seccs is a Python library that realizes a secure and efficient hash-table-like data structure for contents on top of any existing key-value store as provided by, e.g., cloud storage providers.

It has been developed as part of the work [LS17] at CISPA, Saarland University.

Installation

$ pip install seccs

If you want to use AES-SIV encryption (you probably want!), you also need to install PyCrypto 2.7a1 which is not yet available in PyPI:

$ pip install https://ftp.dlitz.net/pub/dlitz/crypto/pycrypto/pycrypto-2.7a1.tar.gz

Usage and Overview

seccs is a Python implementation of sec-cs, a secure and efficient hash-table-like data structure for contents. It stores its data on top of any existing database providing a key-value store interface. Thus, it is likewise usable with in-memory dict objects, persistent databases like ZODB, and many cloud storage providers.

Its details are described in [LS17]. In short, it is suitable for usage on untrusted cloud storage and has the following desirable properties:

• Confidentiality:

  Stored contents are securely encrypted using a symmetric key.

• Authenticity:

  sec-cs guarantees authenticity of all stored contents, irrespective of gurantees of the underlying database.

• Storage Efficiency:

  Data deduplication strategies are applied to all stored contents. When storing new contents, overlapping parts of existing contents are automatically reused as to avoid redundancy. sec-cs is optimized for efficiency in presence of many similar contents: Storage costs of an n-bytes content that differs only slightly from an existing content are in O(log n).

Typical Use Case

In the most-typical configuration, sec-cs chunks its contents hierarchically using ML-CDC (see [LS17]), usually relying on Rabin Karp hashes, and stores the resulting nodes in a database after applying AES-SIV-256 for encryption and authentication. From a user perspective, we have to initialize a suitable database object and a 32-bytes key first.

Database and key setup:

>>> database = dict()
>>> import os
>>> key = os.urandom(32)

Note that we might want to store the database and the key at some persistent location in practice.

Next, we need to create a crypto wrapper which is in charge of all the cryptographic operations. Depending on our security goals (e.g., whether encryption is required), we could choose any suitable wrapper from seccs.crypto_wrapper. Afterwards, we can instantiate the data structure.

Choice of crypto wrapper and instantiation of data structure:

>>> import seccs
>>> crypto_wrapper = seccs.crypto_wrapper.AES_SIV_256(key)  # install PyCrypto>=2.7a1 to use AES-SIV
>>> seccs = seccs.SecCSLite(256, database, crypto_wrapper)  # 256 is the chunk size

Note

Internally, sec-cs splits contents into chunks, creates a tree of chunks for each of them and inserts each node separately into the database. The first parameter specifies the desired average size of nodes inserted into the database. As deduplication is performed at the chunk level, large chunk sizes decrease deduplication performance, but they also create less storage overhead when storing non-deduplicable contents as fewer nodes have to be stored.

Performance is discussed in detail in [LS17]. If high redundancy is expected, 256 bytes is typically a good compromise; otherwise, larger chunk sizes might be more suitable.

We can now insert contents…

>>> content = b"This is a test content."
>>> digest = seccs.put_content(content)
>>> repr(digest)
'\x08,f+\xa74\xdc\x0f\xe5Oo\xcb;\x83\xb9T\x00\x00\x00\x00\x00\x00\x00\x17'

…retrieve them…

>>> seccs.get_content(digest)
This is a test content.

…and delete them as soon as they are not needed anymore:

>>> seccs.delete_content(digest)

Storage Efficiency

seccs avoids redundancy in the database wherever possible, as gets clear in the following example.

Consider this function for measuring the database’s current storage costs in bytes:

>>> import sys
>>> def dbsize(db):
>>>     return sum([sys.getsizeof(k) + sys.getsizeof(v) for (k, v) in db.items()])

Initially, the database is empty:

>>> dbsize(database)
0

Insertion of a 1 MiB content clearly causes some storage costs:

>>> content1 = os.urandom(1024*1024)
>>> digest1 = seccs.put_content(content1)
>>> dbsize(database)
1583030

But inserting the same content for a second time does not incur additional costs:

>>> content2 = content1
>>> digest2 = seccs.put_content(content2)
>>> digest1 == digest2  # identical contents yield identical digests
True
>>> dbsize(database)
1583030

Clearly, the database grows if different contents are inserted. However, these costs are low if inserted contents are similar to existing ones.

Only about 2.3 KiB are required to store another 1 MiB content with one byte changed:

>>> content3 = b''.join([content1[:512*1024], b'x', content1[512*1024+1:]])
>>> digest3 = seccs.put_content(content3)
>>> dbsize(database)
1585395

Costs are similar even if the identical parts are shifted…

>>> content4 = b''.join([content1[:512*1024], b'xyz', content1[512*1024+1:]])
>>> digest4 = seccs.put_content(content4)
>>> dbsize(database)
1588010

…and deduplication is also performed if a content consists of parts of different existing contents:

>>> content5 = b''.join([content1, content3, content4])
>>> digest5 = seccs.put_content(content5)
>>> dbsize(database)
1591009

In the last example, the growth was about 3 KiB.

Furthermore, storage space is reclaimed completely when contents are removed:

>>> seccs.delete_content(digest5)
>>> seccs.delete_content(digest4)
>>> seccs.delete_content(digest3)
>>> seccs.delete_content(digest2)
>>> dbsize(database)
1583030
>>> seccs.delete_content(digest1)
>>> dbsize(database)
0

Note

Every seccs.delete_content call undos eactly one seccs.put_content call. Thus, even if the same content has been inserted twice, yielding only a single digest, it has to be deleted twice as well to get actually removed.

Testing

seccs uses tox for testing, so simply run:

$ tox

References:

[LS17] (1,2,3,4)

Dominik Leibenger and Christoph Sorge (2017). sec-cs: Getting the Most out of Untrusted Cloud Storage. In Proceedings of the 42nd IEEE Conference on Local Computer Networks (LCN 2017), 2017. (Preprint: arXiv:1606.03368)