Slashdot Mirror
New Password Recovery Technique Uses CPU and GPU Together
BaCa writes to mention that a new hardware/software combination has been created by a company called ElcomSoft that will reportedly allow cryptography professionals to build cheap PCs that work like supercomputers for the specific task of retrieving lost passwords. Utilizing a combination of the CPU and the GPU the task of brute forcing a password may be reduced by as much as a factor of 25. "Until recently, graphic cards' GPUs couldn't be used for applications such as password recovery. Older graphics chips could only perform floating-point calculations, and most cryptography algorithms require fixed-point mathematics. Today's chips can process fixed-point calculations. And with as much as 1.5 Gb of onboard video memory and up to 128 processing units, these powerful GPU chips are much more effective than CPUs in performing many of these calculations."
So what, will hackers be able to use my computer to crack my password 25 times faster now?
now IT departments will require passwords to be 30 characters long, with at least 2 digits, at least 2 puncuation marks, mixed case, and use Unicode characters from at least 8 different international languages.
Table-ized A.I.
If brute force isn't working... you aren't using enough of it.
Weeks of coding saves hours of planning.
Oh wait, both.
Even people that believe in pre-destiny look both ways before crossing the street.
Pricing for these apps is pretty steep at $1,299 per machine license. Well, maybe not so steep if you consider how valuable it could be for you. It doesn't say if that has the GPU utilization with it yet or not.
Also, I wonder if they've investigated using SLI & CrossFire with these. That seems like something obvious to me but not included in the article. I'm unaware of their implementation but it sounds like it could be parallelized--and accross 2 or even 4 cards, that could get hilariously powerful.
My work here is dung.
"Password Recovery" sounds so much more benign than "Cracking Passwords".
Hello, Mr. Orwell. *wave*
I can now release the 12,000 monkeys I kidnapped for the task.
Table-ized A.I.
What seems to have been missed in the discussion so far is that this company is applying for a patent on their technique, which they claim is "revolutionary." I really hope that this doesn't get granted, as it would open a whole new realm of stupid patents for "X on a graphics card," which is about as stupid a patent as "X on the internet."
This project has been around for a long time: http://www.gpgpu.org/ Though I agree modern GPU's are even more useful for general purpose computing.
I've read the article (such as it is), and it keeps claiming that this is a technique to recover "lost passwords". But I don't really believe that is the purpose of this software, and I have to ask "What is the difference between a 'lost password" and a password that belongs to someone else and not you?". Does anyone else really believe that the actual use of this software will be to assist the majority of users recover their own passwords? I do not. I suspect it might be harder to patent a tool for identity theft than for recovering "lost passwords" though.
I'm an American. I love this country and the freedoms that we used to have.
FPGAs (Field-Programmable Gate Arrays) sound like they would be just the ticket for SIMD (single-instruction-multiple-data) calculations such as this. Configure up a bunch of FPGA chips to do the encryption calculations on a zillion combinations in parallel...
Paleotechnologist and connoisseur of pretty shiny things.
Petter Nordahl-Hagen's Offline NT Password & Registry Editor: http://home.eunet.no/~pnordahl/ntpasswd/
NOTE: Tested on: NT 3.51, NT 4 (all versions and SPs), Windows 2000 (all versions & SPs), Windows XP (all versions, also SP2), Windows Server 2003 (all SPs), Vindows Vista 32 and 64 bit.
This is the company with which Dmitry Sklyarov was employed at the time of his arrest by the FBI, back in 2001. Before his arrest, at a conference, Dmitry made a presentation on cracking Adobe's eBook DRM. The method used for this crack was utilized in Elcomsoft's Advanced eBook Processor software.
This was really big news back in 2001-2002, although I guess thats a bit too long ago for most slashdot readers, since I (surprisingly) haven't seen any other comments mentioning this.
I'm just wondering, should I take the summary as intentionally ironic (i.e. as if it had referred to an operating system "by a company called Microsoft"), or should I assume it was written by someone *fascinatingly* oblivious to the recent history of decryption software and the disputed legalities thereof? An informed, non-ironic summary would simply say, "...by ElcomSoft...", of course.
For any of you who may have been living under a rock (possibly on another planet), ElcomSoft is the company that was employing Dmitry Sklyarov, who was arrested in the US on DMCA charges when he'd come to present at a conference. Wikipedia has more.
Anyone car to point me to one of these mythical video cards with 128 processors and 1.5 gig of fast on board memory? Also, at the price point they are asking for this software (1200USD per seat) it seems like this is hardly cost competitive with doing this same sort of thing using commercially available FPGA dev/prototype boards and open source software designed for this EXACT task.
-*The above statement is printed entirely on recycled electrons*-
Fail.
...allow cryptology professionals to build affordable PCs that will work like supercomputers when recovering lost passwords. Cut and pasted from "How to write with spin for dummies"Fail.
...will be incorporating this patent-pending technology into their entire family of enterprise password recovery applications. Corporate press release copy and paste == Fail.Numerous grammatical errors == Fail.
Add 1 letter and you've increased the time it takes to hack by 26x (although it's probably closer to 100x with punctuation and the like). So 25x is irrelevant. So is 250x. Only something that makes it non-exponential would really make a difference.
Hello, I would like to order one of your _cheap_ PCs, specifically the one with 128 GPU:s which I will turn into a supercomputer with this great software. I need it to recover my lost windows password. Thank you. And by the way do you still have those low-energy, standard socket 1.21 gigawatt bulbs?
unless you're using a crappy password scheme like Vista's, for example.
This is a process that lets you brute-force passwords 25 times faster. That's pretty neat, I'm not arguing that. It's extremely clever. But this speed [i]shouldn't matter[/i], because cracking passwords a mere 25 times faster shouldn't matter either. The problem comes down to how people are designing a lot of password schemes. They're aiming for speed. The article says the new technique can try ten million passwords per second on a single computer. Division tells us that, beforehand, the computer could process 400,000 passwords per second.
When was the last time you had four hundred thousand users logging into a single computer per second?
Checking a password should be slow. Brutally slow. I mean, quite literally, that just checking to see if the user's password hashes correctly should take at least a hundredth of a second. You're not going to have a hundred users logging in per second on a single computer anyway, our modern database-driven sites couldn't handle the load of displaying the login pages, so why are we making our password schemes so flimsy?
If you use a slow password hash generation - and this can be something as simple as iterating MD5 over itself ten thousand times - whoever's trying to brute-force your password scheme is going to have a horrible, horrible time of it. Add a basic salt to the mix and you will not have anything to worry about from this. If your password checker takes a hundredth of a second, then 25 times faster means your adversary is going to spend $1300 on software in order to try 2500 passwords per second. If you have an appropriate salting system that's 2500 passwords for a single user. This is not the death knell for passwords, or anywhere near it. If anything, it's the death knell for crappy password hashes - but it's not even that, since you could trivially foresee things like this years in advance.
Brute-force password cracking, by its very nature, is millions of times more expensive than merely verifying a valid user. From there, it's up to you to determine how safe you want your passwords to be. Personally? I'm fine with wasting a few extra hundredths of a second per user.
Breaking Into the Industry - A development log about starting a game studio.
GPUs were foremost designed to execute large numbers of linearly-ordered simple matrix/vector operations per clock cycle. When it comes to generating 3D, there isn't much in the way of branching, recursion or conditional execution involved. I haven't checked recently, but it used to be that a "pixel pipeline" referred to a unit that could do a 4x4 * 4x1 operation in a single clock (16 multiplies and 12 adds).
Coincidentally this also helps a large number of scientific applications, such as molecular dynamics, or physics applications that can be converted into vectors and manipulated, such as kinematics (this is what a physics engine often does).
Game tree searches (I've written a few in my time) are usually highly recursive with exponential growth (branching factor). It would be very difficult to transform these into an efficient set of linearly-ordered vector operations. For example the static evaluator on many (older) chess engines consists of a painful set of heuristics and exceptions to heuristics, and exceptions to exceptions to heuristics. It is a very chaotic flow problem.
Education is a better safeguard of liberty than a standing army.
Edward Everett (1794 - 1865)
Cthon98> hey, if you type in your pw, it will show as stars
Cthon98> ********* see!
AzureDiamond> hunter2
AzureDiamond> doesnt look like stars to me
Cthon98> AzureDiamond> *******
Cthon98> thats what I see
AzureDiamond> oh, really?
Cthon98> Absolutely
AzureDiamond> you can go hunter2 my hunter2-ing hunter2
AzureDiamond> haha, does that look funny to you?
Cthon98> lol, yes. See, when YOU type hunter2, it shows to us as *******
AzureDiamond> thats neat, I didnt know IRC did that
Cthon98> yep, no matter how many times you type hunter2, it will show to us as ******
AzureDiamond> awesome!
AzureDiamond> wait, how do you know my pw?
Cthon98> er, I just copy pasted YOUR ******'s and it appears to YOU as hunter2 ause its your pw
AzureDiamond> oh, ok.
Why is the GPU a processor dedicated to nothing but "pretty graphics" so much more powerful than the central multi-purpose processor even at the things like number-crunching?
You need to rephrase your question, because it makes an incorrect assumption. Here:
Why is the GPU a processor dedicated to nothing but "pretty graphics" so much more powerful than the central multi-purpose processor especially at the things like number-crunching?
The answer is obvious if you think about it: those "pretty graphics" are a huge number crunching problem. That's all there is to it. GPU's, however, aren't very good at tasks that don't do exactly the same thing huge numbers of times. This is true of most applications. Including the applications that run on the PC to control what the GPU does in stuff like what the story's talking about.
Is it because the GPU engineers can completely redo the thing from scratch whenever they want to, whereas the CPU-designers are held back by the backwards-compatibility issues?
Partially. Modern GPUs have (I think -- I don't keep up to date) 256 bit wide memory interfaces, running at close to gigahertz speed. This means they can transfer to and from their memory at about 4 times the rate a PC can. This is possible because (1) graphics card manufacturers don't mind the types of memory they use changing on a virtually model-by-model basis and (2) they also don't mind being stuck with non-expandable memory that's soldered directly onto the card right next to the GPU.
It's also because GPU engineers can sacrifice a lot of the flexibility of a PC. So what if the pipeline stalls if all 32 threads aren't doing exactly the same thing at the same time? Most of the time, they will be.
Computer Science teaches, programmers aren't supposed to have to do "tricks" like this -- you code, and the translator (compiler or intepreter) will translate from your programming language to the hardware instructions.
So why did my CS course have a module where we learned how the hardware worked? About memory hierarchies? About SISD, SIMD and MIMD processors? Why does Knuth's The Art of Computer Programming, possibly the most important book ever written on CS, approach problems at an assembly language level? Why, in my CS course, did I learn two different kinds of assembly language (one CISC, one RISC)?
Because CS is concerned with a holistic view of computers. With the fact that they are machines for executing instructions, and what can be done with those instructions. With the fact that it may be more efficient not to specify that much detail, but also the fact that, from time to time, you do need to do that.
a company called ElcomSoft
How short Slashdotters' memories are. ElcomSoft is the Russian company Dmitry Sklarov was working for when he wrote the ebook software that got him thrown in jail when he visited the USA, after Adobe made a DMCA complaint against him.
Minor correction - I know what you mean when you say "linearly-ordered" but a more accurate way to describe it would be: large sets of independent operations per clock-cycle. The sequential encoding that happens between clock cycles is true of most processors, and not specific to GPUs. The key is high performance is the lack of communication between separate instances of the pixel shader which is a property of the independence of the sub-problems.
You're right about the chess search, in contrast to what the poster above you claims, a GPU would not be suitable for evaluating the heuristic because of the branching control flow within the heuristic. An interesting scoring function would be to try and encode a neural network that can score the position: then data packing would be an issue but a neural net can be converted into very efficient GPU code as long as the number of pixels that you are gathering the position from is quite low.
The real performance killer would be the scatter at the other end - once a position is scored the card needs to perform sorting to filter out the high scoring positions from the low. This sorting may be avoidable on the newer CUDA cards as their memory architecture allows an efficient scatter without the need to sort data.
Slashdot: where don knuth is an idiot because he cant grasp the awesome power of php
Anyone ever considered using the PS3 for stuff like this? Seems like you have all the processing power you need (relatively speaking), but what else would you need to take into consideration?