Efficient binary-to-string formatting (like base64, but for UTF8/UTF16)?

Question

I have many bunches of binary data, ranging from 16 to 4096 bytes, which need to be stored to a database and which should be easily comparable as a unit (e.g. two bunches of data batch only if the lengths match and all bytes match). Strings are nice for that, but converting binary data blindly to a string is apt to cause problems due to character encoding/reinterpretation issues.

Base64 was a common method for storing strings in an era when 7-bit ASCII was the norm; its 33% space penalty was a little annoying, but not horrible. Unfortunately, if one is using UTF-16, the space penalty is 166% (8 bytes to store 3) which seems pretty icky.

Is there any common storage method for storing binary data in a valid Unicode string which will allow better efficiency in UTF-16 (and hopefully not be too horrible in UTF-8)? A base-32768 coding would store 240 bits in sixteen characters, which would take 32 bytes of UTF-16 or 48 bytes of UTF-8. By comparison, base64 coding would use 40 characters, which would take 80 bytes of UTF-16 or 40 bytes of UTF-8. An approach which was designed to take the same space in UTF-8 or UTF-16 might store 48 bits in three characters that would take eight bytes in either UTF-8 or UTF-16, thus storing 240 bits in 40 bytes of either UTF-8 or UTF-16.

Are there any standards for anything like that?

Answer 1

Base32768 does exactly what you wanted. Sorry it took five years to exist.

Usage (this is JavaScript, although porting the base32768 module to another programming language is eminently practical):

var base32768 = require("base32768");

var buf = new Buffer("d41d8cd98f00b204e9800998ecf842", "hex"); // 15 bytes

var str = base32768.encode(buf); 
console.log(str); // "迎裶垠⢀䳬Ɇ垙鸂", 8 code points

var buf2 = base32768.decode(str);
console.log(buf.equals(buf2)); // true

Base32768 selects 32,768 characters from the Basic Multilingual Plane. Each character takes 2 bytes when represented as UTF-16 or 3 bytes when represented as UTF-8, giving exactly the efficiency characteristics you describe: 240 bits can be stored in 16 characters i.e. 32 bytes of UTF-16 or 48 bytes of UTF-8. (Except for the occasional padding character, analogous to the = padding seen in Base64.)

This is done by dicing the input bytes (i.e. 8-bit unsigned numbers) into 15-bit unsigned numbers and assigning each resulting 15-bit number to one of the 32,768 characters.

Note that the characters chosen are also "safe" - no whitespace, control characters, combining diacritics or susceptibility to normalization corruption.