zxcvbn compatibility

[gosimplellc]

zxcvbn compatibility

[Tostino]

Hi there @TomLottermann, that observation it wasn't an overriding goal of nbvcxz to maintain complete compatibility with zxcvbn or the other ports is entirely correct.

When I started this project, there were no other Java ports and I really wanted this functionality for my company, and I wanted it in Java. By the time I was done with the initial version, zxcvbn4j was also released, and from my memory of looking over their code years ago, they did maintain closer compatibility with zxcvbn (at least at the time).

Now on to ways nbvcxz can be configured to generate closer output to zxcvbn:

1. Disable the Levenshtein Distance (LD) calculation. This feature was very helpful in my analysis on helping identify passwords which were only slightly different than dictionary words but were not caught with the original implementation. This feature will be sure to cause nbvcxz to produce different results than zxcvbn for a large number of passwords.

2. Make sure both implementations are using the same dictionaries. There are many additional leaked passwords in the nbvcxz dictionary than in zxcvbn. There are also additional dictionaries included in nbvcxz that are not in zxcvbn and vice versa. Simply different choices on what lists were important to include by default. With nbvcxz you can easily change what dictionaries are being used though, so it's easy to make the different implementations use the same dictionaries.

3. The algorithm to find the best matches is different between nbvcxz and zxcvbn, that is likely to produce slightly different results in cases where zxcvbn is unable to find the best combination of matches due to the algorithm used. There were quite a few instances I noted that brought about the change to the algorithm used by nbvcxz where there were obviously "wrong" results for entropy based on the combination of matches because it got stuck in a local minimum. This is no longer an issue with nbvcxz, but will inherently produce different results for some passwords compared to the original algorithm used by zxcvbn. In the majority of cases both algorithms are able to figure out what the lowest entropy combination of matches on the password are, so I don't see this being too big of an issue.

Hope that helps, and i'm interested in your findings if you end up testing nbvcxz and another implementation to see how similar their outputs are over a sample of passwords.

[Tostino]

Another difference I thought of today is the separator match type support we have. It helps with passphrases detection a lot, but since zxcvbn doesn't support it, that would be something to also disable using the ConfigurationBuilder: setPasswordMatchers(List passwordMatchers).

[Tostino]

@TomLottermann I am going to close this issue, hopefully my answers were helpful. If you have any more questions or need anything clarified feel free to ask.

[TomLottermann]

Sorry for the late reply.

Thanks a ton for the detailed info! Might we worth documenting this in the README - maybe :)

[Tostino]

Good call, i'll re-open this as a reminder to add this info to the README.md under a "compatibility" section.

[Tostino]

@TomLottermann I updated the readme with the info from this thread, let me know if you think anything is unclear or should be reworded if you wouldn't mind.

[TomLottermann]

Thanks for the documentation update! :)

[thomas-mc-work]

Is there a shorthand to simply remove the SeparatorMatcher from the existing list of default matchers within a configuration or during the configuration building process?

[Tostino]

Nope, no shorthand that I am aware of.