[tptacek]

No, this is retarded; the "unique nonreversable visualization" is trivially attacked offline by brute forcing with a dictionary of passwords. It's just the SHA1 of your password; fundamentally no different from tacking /etc/shadow to your password field as you type.

It wouldn't be retarded if it wasn't expending a lot of effort (and confusing the hell out of users) to get to that bad destination. But that's what it does.

When you consider "solutions" to this "problem", model it against an adversary with a camera.

The author of this post means well, but just masking the password characters, like every secure system has done for the past couple decades, remains the right answer.

[umbrae]

Well, as the author, I feel I better defend myself here.

A: It only uses the first 20 bytes of the hash. You could narrow it down from this if you were really determined, but you'd not be able to reverse it.

B: The visualization of the sparkline doesn't have the fidelity to determine between characters 6 and 7. So you'd have a range of possible characters.

C: The alternative being suggested by Jakob Nielsen is no masking at all ( http://www.useit.com/alertbox/passwords.html ) - which is less secure? I know this isn't the best argument, but it still is -an- argument.

With that out of the way, my paranoid mind agrees with you in this context: just masking the passwords is the more secure solution. But that doesn't mean that experiments to provide a more usable approach with (arguably) equal security should be avoided.

[tptacek]

(a) This sentence doesn't make any sense. You can't "reverse" full SHA1 any more than you can "reverse" truncated SHA1. And SHA1 hashes are only 20 bytes long. Reversing isn't the attack.

(b) The goal of the attack isn't to magically conjure the password; it's to magically conjure a searchlist of several tens of passwords, which is a game-changing improvement over a searchlist of, say, 72^8 passwords, or even tens of thousands of dictionary words.

(c) The alternative suggested by Jakob Nielson is manifestly and categorically asinine.

Good on you for a finding an application for visualizing a SHA1 hash. You score maximum points for cleverness. But now you should retire this idea.

[umbrae]

(a) I wrote in haste. Obviously it's a one way hash and unreversible. Additionally, I meant the first 20 characters, which is half of the full hash.

(b) Obviously. Still extraordinarily difficult given this implementation IMO. But I take your point genuinely.

(c) Agreed.

If you weren't so friggin inflammatory I'd think we could come to a conclusion here. I could definitely be using a more lossy visualization to be more secure. This is something I'll look into, even though I'm sure you'll still consider it 'retarded' even if it helps your grandma login to her googles more often, making you get less phone calls to fix it as a result. She's got that palsy you know.

[tptacek]

You think running a dictionary through SHA1, generating little sparkline graphs for each hash (the same way you did in like 10 lines of JS code), and then carving the image into sectors and counting matches is "extraordinarily difficult"? Everything you need to do it is, I think, in Mochikit.

You get full props for grandma, her googles, and the palsy, although anything you could do to allow either of my grandmothers to log into the googles would qualify you for much more than HN props. I'm sorry you think I'm inflammatory, except you and I both know I'm not.

[paulgb]

If the attacker had a video camera and the victim was a slow enough typer (I'm thinking hunt-and-pecker), brute force wouldn't even be necessary. You could incrementally build up a string that matched the password as each letter was typed using the hash generated after each letter is typed.

Fidelity of the graphs wouldn't matter at this point, you would just take the closest match and backtrack if necessary.

[tptacek]

Oh wow, I didn't even notice that it generated graphs on a keystroke/timer (though, I mean, duh). Wow, is this bad. Number of mask dots, 2 midpoint colors, and the final color; how many passwords in any dictionary does that reduce to?

[lhorie]

>> model it against an adversary with a camera.

Which can record your hands? :)

I figure that the offline dictionary attack could be foiled if this was a Firefox extension that generated a random salt on installation. (of course, this doesn't work if you want to play WOW on an Internet Cafe)

My main reaction to the experiment is that I don't know many people who touch type (at least when it comes to their password): I've had people accidentally type their password on the username field in front of me countless times because they weren't even looking at the screen.

I don't see why I should worry about big brother FUD when I could embarrass myself any time by accidentally pressing caps lock instead of tab.

[tptacek]

Another thing I love about this discussion is the intimation ("big brother") that in order to get a picture of my screen, you have to be the NSA. And not, you know, some jackass with a camera.

[lhorie]

Sure, sure, it could be both. It could be my girlfriend looking over my shoulder too. But of course, I don't use an stupidly easy password and I don't let my girlfriend see me typing it (2-finger typing style, of course). I mean who does that, right? </sarcasm> :)

I just find it interesting to observe the discrepancies between perception of security and actual math-backed security.

[tptacek]

I don't use a stupidly easy password either, but I'm not going to give you the password hash from my laptop. That's what this scheme does.

[cstejerean]

How often do you type your password next to an adversary with a camera?

[tptacek]

I did it about 50 times today. I'm a building full of cameras. Tell me, how often do you sweep the surrounding area for cameras before you log in somewhere? Thought so. How many times have you ever logged into something from an airport? Ever hear about the Microsoft Blue Hat talks about the pen testers who collect credentials at Sea-Tac?

People just like memes. They're fun to talk about. I promise you there are 1,000 better problems for you to tackle in your app than the suboptimality of password masking. This guy managed not only to waste time, but also to promote an actively evil security extension.

[sho]

Why are you worried about how people with cameras could eventually brute force your password from that tiny, fuzzy smear of colour on the screen, when those same cameras can see your fingers typing the whole damn thing in, in plaintext?

[tptacek]

Because those cameras can see my screen easily, and not my fingers typing on the screen? Because my fingers aren't perpendicular to the floor? Because even if I was going to concede that the image of the SHA graph was going to be fuzzy, just getting the vicinity of the color reduces my search space by 1/2^(16..24)?

I love "cameras could eventually brute force" in your sentence, as if we were actually talking about some crazy hard feat of image analysis and cryptography, rather than running a dictionary through SHA1 and comparing colors.

Got any more arguments, sho?

[sho]

If someone is trying to learn your password from cameras nearby, they will make sure they can see your keyboard. Of course. If you're in that kind of hostile environment, nothing is safe except biometric or card authentication. Your concern is misplaced.

You should have raised the scenario of, say, spying on a screen through a window from the other side of the street. That's a much better example and in that one, you might actually have a point, although I understand it's against the rules in many organisations to allow a screen to face a window. But the "in the same building" one was ridiculous. Why bother brute forcing your squiggly coloured blotch when they can just put a tiny camera in every light, or whatever?

No, no further arguments. I agree it's a bad idea, of course.

[dfranke]

There are plenty of scenarios where someone would be watching you type your password through a camera, but doesn't control placement of the camera. The most common case would be a security guard who is paid to monitor the cameras but had no hand in the design of the surveillance system.

[thaumaturgy]

The goal regardless is to not provide a visual clue to an eavesdropper which they could use discover your password.

I actually don't think the idea of this is all that terrible, just that the implementation isn't right. For example, I can't think of a reason why the representation needs to be unique -- the user would only need a clue if their password is the wrong length or if they likely have a typo. Also, the specific representation used is too resource-intensive; three colored boxes would work just as well.

Still, even if the system returned hundreds of possible passwords for a given hash, and even if it wasn't resource-intensive, it's still providing clues to a third party, which you don't generally want to do. You're still dramatically decreasing their search space.

EDIT: On second thought, there might be a way to make this a little less horrible. If you assume that the user will generally be very close to correct in typing their password, then you could tune a function that would produce dramatically different results for similar passwords, and similar results for dramatically different passwords -- a kind of inverted hash function. Such a function probably already exists, I'm just not familiar with it. If you did that, and completely obscured what the user was typing, so that an eavesdropper couldn't tell the length of the password, then you might be increasing the search space compared to simply displaying asterisks.