Quote from: wyager on January 31, 2014, 12:09:09 AM

Bump. Would like to hear opinions on bloom filter.

The Mycelium Wallet on Android already uses scrypt for its password-protected backups, and for BIP38. On very old devices we have trouble running this, so we will have to introduce a "weaker" version for devices with < 20MB per process.

For a future standard for HD wallet, non-hardcoded scrypt parameters would be preferable, depending on password length and device speed and usage frequency, and usecases (offline device) the desired parameters could be different. I like scrypt, though. If it turns out that scrypt is broken, there should be a V2 of this spec, with a new hashing algo, or the ability to specify the KDF algo.

Some comments on what i have read so far:

1) direct encoding of scrypt parameters would be preferable as direct values as opposed to versioned parameter lists (up& downwards compatible)

2) The generation date is not really that important. Maybe other wallets need it but out software would ignore it with the current scheme, but i see how it potentially speeds up initial syncing for a fresh backup. but thinking long-term it will not help much, since the time between backup and restore will always increase.

3) If the KDF is offloadable, make sure that the "non-offloadable" part does not strictly require a scrypt implementation, so that it may run on devices with very limited resources. (think future revisions of embedded devices)

4) We absolutely want the checksum with bloom filtering in future, so that the user may specify additional passwords/pins for multiple identities. Our non-trivial job will be to design a wallet format that preserves the deniability with the stored metadata.
again, the chosen parameters should not be hard-coded, but i can imagine that default values like 4 bytes with 6 hash functions (optimum for N=4, p=0.02) would make sense.
the question is, do we want to push the plausible deniability so far, as to obfuscate the fact how many additional passwords possibly exists? if so, we would hit a limit after adding 4 passwords,

5) what i do not fully understand, is how bloom filtering impacts the preH entropy. can you elaborate on this?