Block cipher

The block cipher process 64-bit blocks in 56 rounds. It uses 8 bits from an expanded key on each round.

Stream cipher

The first 32 round of the stream cipher are used for initialization and do not generate any output. The first 64 bits of data are used as initialization vector during this phase and are left unchanged. The stream cipher then generates 2 bits of pseudo-random stream on each round which are xored starting at bit 64 of the packet.

Software implementations and bit slicing

The stream cipher part of CSA is prone to bit slicing, a software implementation technique which allows decryption of many blocks, or the same block with many different keys, at the same time. This significantly speeds up a brute force search implemented in software, although the factor is too low for a practical real-time attack.

The block cipher part is harder to bit slice, as the S-boxes involved are too large (8x8) to be efficiently implemented using logical operations, a prerequisite for bit slicing to be more efficient than a regular implementation. However, as all operations are on 8-bit subblocks, the algorithm can be implemented using regular SIMD, or a form of “byteslicing”. As most SIMD instruction sets, (with the exception of AVX2) do not support parallel look-up tables, the S-box lookups are done in a non-bytesliced implementation, but their integration into the rest of the algorithm is not hampered markedly by the byteslicing.

Both techniques are used in libdvbcsa, a free implementation of CSA.

Cryptanalysis

Cryptanalysis is made more difficult as most data is protected both by the block and the stream cipher. Adaption codes can result in packets with some data protected by only the block cipher or the stream cipher.^[2]

Brute force approach

While the CSA algorithm uses 64-bit keys, most of the time, only 48 bits of the key are unknown, since bytes 3 and 7 are used as parity bytes in CA systems, and may be easily recalculated. This allows for possible known-plaintext attacks when combined with knowledge of the underlying plaintext structure. As the first three bytes of the PES header are always 0x000001, it would be possible to launch a brute force attack. Such an attack would reveal millions of possible keys, but still few enough to make it practical to attempt decryption of other parts of the data with the same key in a second pass to recover the true key.

Known-plaintext attack

In 2011, a group of German researchers released an attack^[3] on CSA as used in the DVB system. By noting that MPEG-2 padding frequently requires long series of zeroes, leading to entire 184-byte cells being encrypted with zeroes only, it is possible to build up a rainbow table recovering the key from such a known-zero block. (A block would be known to be zero if two blocks with the same ciphertext were found, since presumably both would be zero blocks.) The attack described would require about 7.9 TB of storage, and enable an attacker with a GPU to recover a key in about seven seconds with 96.8% certainty. However, the attack is only effective when such all-zero padding blocks are present (i.e., in scenes with very little movement or noise), and it would be easily mitigated by for instance sending such all-zero cells unencrypted.

This differs from the brute-force attack examples already mentioned, in that the plain-text contents of the entire cell is known before the attack, not just parts of a block, which enables the ahead-of-time computation of the rainbow table.

Fault attack

In 2004, a fault attack^[4] was published on the block cipher. The basic idea was to introduce a few deliberate faults in the intermediate calculations, making it possible to deduce the last eight round keys. From this, the final key can be computed.

While very few faults are needed (on average two or three per round key recovered), fault attacks are usually impractical to carry out, in that it requires the ability to change bits at will inside the hardware of a decoder that already has the key in question.