There's an easy way to measure entropy non-heuristically: see how well the password compresses. Then, to measure whether a new password is a variation of an old one, you just compress their concatenation and see if you get less entropy than expected.

 That requires storing the old password in plain text (or something reversible to plain text, like lossless compressed formats). That is typically not recommended as old passwords can give clues to the current password.
 Nah; you can require that they enter the old password to change their password, check that they've entered the right one, and then use that (while you're still holding onto it) to do the cross-compression check.
 If you have both the old and new password in plain text you don't need to do a compression-based check, just a raw iterative approach would accomplish the same thing with very low overhead (in either CPU or memory accesses).Compression systems typically need a lookup table of some kind and have more overhead than just the raw comparison.
 As mentioned in another comment, I'd probably do a Levenshtein distance between the old and new passwords, and reject if they crossed some threshold. However, only knowing the plaintext of the immediately-preceding password as they enter it to authorize the change, it wouldn't do much to stop them from doing:PasswordA SomeOtherPassword1 PasswordB SomeOtherPassword2 PasswordC SomeOtherPassword3Just iterate on every other change, and you've beaten the requirement.
 That's assuming what you want is to find passwords that are "the same" in a literal sense, rather than "the same" as in sharing a common generation algorithm that biases certain outputs.For example, if a user's first password is "first123!@#" and their second password is "second456\$%^", there are no letters in common—but those two passwords, when joined together, are very intracompressible (by an ideal compressor)—and likewise, an attacker who knew that the first was a previously-used password would be very likely to try the second. That both properties apply is not a coincidence of this particular password; the intra-compressibility of a set of plaintexts, and the predictability of unknown plaintexts of the set from known ones, are equivalent measures of informational entropy.

Search: