>The minute you publish your method, the measure of the output's entropy stops being relevant. Your security theoretically relies on the input's entropy, which is likely to be much smaller.
What? If we're applying Kerckhoff's principle, then the mechanism literally never generates entropy. The only place it can ever come from is the source, which means the only thing your methods can do is either keep the same amount of entropy or reduce it (anything else would require another source of entropy). This comes straight from the math: self-information can't increase using mappings alone, but it sure as hell can decrease. It's these decreases that are difficult to measure.
>He actually has to model the desired target's strategy, and produce an attack suited to that.
In theory, yes, but in practice, machine learning does that for you. The point is, a patterned method like using code will require very long passwords compared to just using diceware (or hell, just using /dev/random to pick words from your dictionary file) if you want them to both have the same amount of entropy. The only other argument that can be made to the contrary is that something other than entropy determines password strength, which isn't true if we assume a sufficiently advanced adversary (of which I consider machine learning algorithms + a gpu/asic cluster to be). Since the only things that really go into how "good" a password is is usability and security, and diceware passwords are probably just as or more memorable than code snippets, again, I don't see the advantage of the latter in any case.
What? If we're applying Kerckhoff's principle, then the mechanism literally never generates entropy. The only place it can ever come from is the source, which means the only thing your methods can do is either keep the same amount of entropy or reduce it (anything else would require another source of entropy). This comes straight from the math: self-information can't increase using mappings alone, but it sure as hell can decrease. It's these decreases that are difficult to measure.
>He actually has to model the desired target's strategy, and produce an attack suited to that.
In theory, yes, but in practice, machine learning does that for you. The point is, a patterned method like using code will require very long passwords compared to just using diceware (or hell, just using /dev/random to pick words from your dictionary file) if you want them to both have the same amount of entropy. The only other argument that can be made to the contrary is that something other than entropy determines password strength, which isn't true if we assume a sufficiently advanced adversary (of which I consider machine learning algorithms + a gpu/asic cluster to be). Since the only things that really go into how "good" a password is is usability and security, and diceware passwords are probably just as or more memorable than code snippets, again, I don't see the advantage of the latter in any case.