1 unstable release
0.1.0 | Mar 22, 2024 |
---|
#2433 in Encoding
36KB
377 lines
๐ฌ Bunk.
Fast and efficient human-readable data encoding!
Bunk encodes binary data as pronounceable gibberish, somewhat resembling Latin. This is useful when binary data such as an encryption key is shown to an end-user who might need to manually transfer it.
Using the default settings, a string of 32 bytes gets encoded as:
atemorni telphocom neideu gepypi forzamar oasal cevanal butthepo aujoate turviy menkais
Optionally, Bunk can decorate the encoded string with commas, periods, and sentence casing to improve readability:
Atemorni telphocom. Neideu gepypi forzamar oasal cevanal butthepo aujoate turviy, menkais.
๐ Documentation
Read the documentation HERE.
๐ Overview
- It is fast! On my machine, encoding and then decoding a random array of 32 bytes takes an average of ~0.8ยตs with the default settings โ allocations and all; no hidden fees.
- It is small! Bunk stores a table of only 256 syllables, each between 1-4 letters (average of 2.47), and some data structures needed for fast lookup.
- Checksums of variable length can be added to encoded messages to verify data integrity when decoding, which protects against typos.
- The maximum word length (in syllables) can be customized.
โ๏ธ How it compares to English dictionary encodings
A popular scheme is to encode binary data as actual English words, which yields results that are more readable and easier to remember. See bip39 as an example of this. However, to be efficient in the amount of data a string of words can encode, a massive table of (sometimes quite long) words must be included โ bip39 uses 2048 words. In addition to this, some kind of data structure for lookup is also needed, and will likely have to be constructed at runtime. If this is of no object to your application, use something like bip39 instead!
Bunk takes a different approach, requiring a table of only 256 1-4 letter syllables, each carrying one byte of data. This allows Bunk to:
- Take up less memory overall.
- Store data structures needed for fast lookup in static memory instead of having to construct it at runtime.
๐ ๏ธ How it works
To explain the algorithm, we'll iteratively build upon it and solve issues as we go.
The fundamental idea is to encode a byte as a syllable by using it to index into a table of 256 unique syllables, the result of which is then appended to the encoded string โ as one would expect. The decoder can then use a trie to find the index of the longest syllable at the beginning of the string, which corresponds to the encoded byte.
This by itself causes issues of parser ambiguity when one valid syllable is a prefix of another. Take as a basic example the encoded string "ous". Is this the single syllable "ous", or the syllable "o" followed by "us"? Barring some cumbersome machinery, there is no way for the decoder to know! The encoder therefore has to detect when such an ambiguity is possible by checking if the first letter of the second syllable is a valid continuation of the first syllable. If so, it inserts a word break between them. (Technically, this is stricter than necessary for breaking the ambiguity but is easy to check and allows the decoder to be written greedily.)
To support these two required operations โ finding the longest syllable prefixed to a string, and checking whether a letter is a valid continuation of a syllable โ Bunk uses a trie. There are then two issues presenting themselves:
- Tries are slow to construct.
- There are (somehow) no efficient trie libraries for Rust that allows for these operations in their API.
As a solution to both of these, a precomputed trie (as created by crawdad) is stored in static memory, on top of which Bunk implements a basic traversal, which the only API needed for the two operations. All in all, the trie API comes out to only about 60 lines of code โ much less than having to add crawdad (or such) as a dependency.
So far, the algorithm we've described is a perfectly functional encoder. However, to be more user-friendly,
we'd ideally also like all inputs to yield equally pronounceable text. Without any further measures, inputs
such as [0, 0, 0, 0]
yield repeated syllables, in this case "uuu u". To avoid this, Bunk artificially
increases the apparent entropy of encoded bytes by first XORing them with a value dependent on their index.
Since XOR undoes itself, the decoder can then do the exact same thing and retrieve the original bytes. With
this in place, [0, 0, 0, 0]
gets nicely encoded as "trirori mulry".
Dependencies
~240โ760KB
~18K SLoC