Slashdot Mirror
Hydan: Steganography in Executables
An anonymous reader says "Ever wanted to hide a message into an executable? Now you can with Hydan. Presented recently by Rakan El-Khalil at Defcon and Blackhat, this tool lets you embed data into an application without changing its functionality or filesize! Check it out. Use includes steganography as well as embedding a program's signature into itself to verify it's not been tampered with."
"What are you doing?"
"Oh, hydan out."
Discovered a copyright string that was also executeable 68k code .... and included it in my main initialization routines
I am 1337.
"That's the sort of blinkered, philistine pig ignorance I've come to expect from you non-creative garbage."-Monty Python
if you blurt something like that out in the blurb maybe it would be nice to mention how the hell it happens. especially when the site gets slashed so fast.
executable packing or actually increasing the filesize? either one has to happen.
world was created 5 seconds before this post as it is.
it looks like the information is being hidden by a slashdotted executable.
especially if the OS goes off and double checks the executable is legit before executing it...
Donald 'Duck' Dunn: We had a band powerful enough to turn goat piss into gasoline.
If you embed a signiature of the file into the file, this by definition changes the file's signiature. At best you can append the signiature. However if the file can be modified, so can it's signiature.
If these folks have figured out a way of circumventing this innate paradox, I'm impressed and am dying to hear more about the technology/mathematics behind it! Can you say Nobel Prize nomination?
At least not without a top down Orwellian soceity where all hardware and software is controlled.
Free Mac Mini Yeah, it's
Not really :)
But I'd like to make that dog downstairs stop barking.
Get your own free personal location tracker
without changing file sizes... let me stick my pirated version of War and Piece in my Hello world application.
sometimes you don't even have to rtfa to rip on a topic...
Nuttles
Christian and proud of it
Not only a dupe, but a link to the original story is listed on the referenced page.
Wow.
The message retrieval method should be called "Hydan Seek"
Hydan: Hiding Information in Program Binaries
Rakan El-Khalil and Angelos D. Keromytis
Department of Computer Science, Columbia University in the City of New York
{rfe3,angelos}@cs.columbia.edu
Abstract. We present a scheme to steganographically embed information in x86
program binaries. We define sets of functionally-equivalent instructions, and use
a key-derived selection process to encode information in machine code by using
the appropriate instructions from each set. Such a scheme can be used to watermark
(or fingerprint) code, sign executables, or simply create a covert communication
channel. We experimentally measure the capacity of the covert channel by
determining the distribution of equivalent instructions in several popular operating
system distributions. Our analysis shows that we can embed only a limited
amount of information in each executable (approximately 1
110 bit encoding rate),
although this amount is sufficient for some of the potential applications mentioned.
We conclude by discussing potential improvements to the capacity of the
channel and other future work.
1 Introduction
Traditional information-hiding techniques encode ancillary information inside data such
as still images, video, or audio. They typically do so in a way that an observer does not
notice them, by using redundant bits in the medium. The definition of "redundancy"
depends on the medium under consideration (cover medium). Because of their invasive
nature, information-hiding systems are often easy to detect, although considerable work
has gone into hiding any patterns [1]. In modern steganography, a secret key is used to
both encrypt the information-to-be-encoded and select a subset of the redundant bits
to be used for the encoding process. The goal is to make it difficult for an attacker to
detect the presence of secret information. This is practical only if the cover medium has
a large enough capacity that, even ignoring a significant number of redundant bits, we
can still encode enough useful information.
Aside from its use in secret communications, an information-hiding process [2] can
be used for watermarking and fingerprinting, whereby information describing properties
of the data (e.g., its source, the user that purchased it, access control information,
etc.) is encoded in the data itself. The "secret" information is encoded in such a manner
that removing it is intended to damage the data and render it unusable (e.g., introduce
noise to an audio track), with various degrees of success.
In this paper, we describe the application of information-hiding techniques to arbitrary
program binaries. Using our system, named Hydan, we can embed information
using functionally-equivalent instructions (i.e., i386 machine code instructions). To determine
the available capacity, we analyze the binaries of several operating system distributions
(OpenBSD 3.4, FreeBSD 4.4, NetBSD 1.6.1, Red Hat Linux 9, andWindows
XP Professional). Our tests show that the available capacity, given the sets of equivalent
instructions we currently use, is approximately 1
110 bits (i.e., we can encode 1 bit
of information for every 110 bits of program code). Note that we make a distinction
between the overall program size and the code size. The overall program size includes
various data, relocation, and BSS sections, in addition to the code sections. Experimentally,
we have found that the code sections take up 75% of the total size of executables,
on average. For example, a 210KB statically linked executable contains about 158KB
of code, in which we can embed 1.44KB (11, 766 bits) of data.
In comparison, other tools such as Outguess [1] are able to achieve a 1
17 bit encoding
rate in images, and are thus better suited for covert communications, where data-rate
is an important consideration. The 1
110 encoding rate achieved by the currently implemented
version of Hydan is obtained when we only use instruction
If I transmit files out to my friends that include encrypted data using steganography, then the extra data should be indistinguishable, effectively hiding within the noise of random crap on the web/usenet/email. Thus, without the key, an intercepted message is difficult to detect, and even if detected, I have sufficient plausible deniability to say "nothing there".
In order to detect an message encrypted and included inside another file, you either need to know its there and be looking for it, compare it to an existing file which should be identical, or statistically detect some aspect of the file. If you know it should be there, you just need to grab any file that looks like the file you're seeking, grab the relevant bits, and attempt decryption. If you have a file that should be identical, (say, an image that looks the same that was posted to usenet a couple days earlier), you can take the bits that are different and try and make some sense of them. If you are just doing statistical analysis, you might be able to find files which have a set of bits whose randomness is just shy of where it should be, and maybe those bits mean something.
In short, unencrypted steganography isn't particularly useful, but encrypted, you can really hide things.
Intresting. Allthough I didn't get a chance to RTFA, hiding encrypted data in an executable doees not seem all that practical to me. It may not change the filesize or functionality, but would it not also change other signature methods (like md5sums?). From my understanding, the main strength of steganography is the file with the encrypted data being indistinguishable from regular files. Since the diffrence can be detected with CRC or MD5, wouldn't that defeat the main purpose?
They should have put their message in the web servers executable so that when it gets slashdotted it could just shit itself and we could still get how it works.
The gist of it is that there are many instructions in x86 that have the same result. You can replace these, and based on which instructions you encounter you can find a hidden message.
So much for theory. Here's an example; let's say we have a couple of synonyms, like so
car, automobile; Robert, Bob; crashed, trashed; beer, whisky.
Let's say we have a little story like so;
"Bob got in his car. He crashed it, because he had been drinking too much beer. His car is now a total loss."
Let's say we want to send a secret binary message "0110". Cunningly, we substitute the first of each pair of synonyms if we want to encode a zero, and the second for a one. So the story is now
"Robert got in his automobile. He trashed it, because he had been drinking too much whisky. His car is now a total loss." (notice how not all key words changed).
This is a bit harder with natural language, as many words aren't quite right to use in place of the other ("got in his automobile" just doesn't sound right), so it's actually easier to do for machine code.
SCO employee? Check out the bounty
inc ax
add ax, 1
add al, 1
inc eax
add eax, 1
All of these i386 instructions do the same thing, but they've got different binary representations. If you encode your information by which instruction you use, you can hide the message without changing filesize or functionality.
"They redundantly repeated themselves over and over again incessantly without end ad infinitum" -- ibid.
OK, you place the hash in the executable; the file is changed. Now the hash should be different...
Problem.
If you mod this up, your slashdot background will turn into a beautiful sunset!
The site is slashdotted so I don't know if this is how it works, but...
Some 8086 opcodes contain a bit that reverses the operation. For examble, with the bit set in the instruction "mov bx,cx", bx would be copied to cx instead of cx to bx. By switching the registers AND setting the bit, you effectively reverse the operation twice, creating different machine code that does exactly the same thing.
The A86 assembler used this bit to create a fingerprint that would make it easier to detect non-paying users.
Comment removed based on user account deletion
If the program has been tampered with, the most obvious thing to tamper with would be the validation mechanism.
I'm going to stick with a separate md5sum, thanks.
-- perl -e'print pack"H*","6e656d6f406d38792e6f7267"'
Suppose eax = 0xFFFFFFFF.
Result of "inc ax": eax = 0xFFFF0000
Result of "inc al": eax = 0xFFFFFF00
Result of "inc eax": eax = 0x00000000
They don't do anything near the same thing. The carry bits get lost.
However, you can substitute "add ax, 1" for "inc ax", and "add al, 1" for "inc al", and "add eax, 1" for "inc eax".
Ironically, on the other hand, emacs has been doing this for years, without much data compression.
*ducks*
Tired of legitimate data sources? Try UNCYCLOPEDIA
Since a lot of people is asking, here it goes:
- How it works
--------------
Overview: Hydan finds sets of equivalent instructions in the binary,
and uses that redundancy to embed data. The larger the set of
equivalent instructions, the more bits can be embedded. For example,
if instructions {a, b, c, d} are all equivalent, then we can embed two
bits of information when any of those instructions are encountered.
Embedding: Hydan goes through the application sequentially, and
whenever it finds an instruction that it has equivalents to, it
substitutes in the instruction that represents the bit(s) of data
hydan is currently embedding. A simple example: "add %eax, 50" is
equivalent to "sub %eax, -50". So this set is {"add %reg, $imm", "sub
%reg, $imm"}. Whenever an instruction of the form "add %reg, $imm" is
encountered, hydan can embed one bit of the message. If the bit is 0,
it leaves it as an add instruction. Else it substitutes it to "sub
%reg, -$imm". (and vice versa)
Decoding: When it is time to extract the embedded message, every
"add %reg, $imm" is taken to mean bit 0, and every sub instruction
encodes the bit 1, and the embedded message is reconstructed that way.
Encryption: Hydan first prompts the user for a passphrase before
embedding or decoding the contents of the application. In the case of
embedding, hydan prepends the length of the message to the message,
encrypts that with blowfish in cbc mode, and embeds the result into
the application. When decoding, hydan extracts all the possible bits
from the application (since it does not know how long the message is
a-priori; that information is encrypted). Hydan then decrypts the
message properly since it is in CBC mode and need not know the total
length first. The lenght is then used to truncate the message to
size.
Instructions: For a complete list of the sets of equivalent
instructions, please refer to hdn_insns.c.
- Attacks
---------
There are three classes of attacks that are applicable to hydan:
overwriting, detection, and extraction. The overwriting attack refers
to the ability to overwrite the message embedded in the application,
whether its presence was detected or not. An attacker should also not
be able to detect the presence of a message in the application, nor
decrypt it.
The overwriting attack: hydan currently has no means to protect
against this type of attack. Since hydan embeds the message
sequentially, starting from the top of the application, an attacker
could re-run hydan with a bogus text and embed that on top of the
original message. The intended recipient of the application would
thus be unable to retrieve the original message. One way this could
be solved is to add an error correcting code to the encoding of the
message, and distribute the message throughout the application in a
passphrase specific manner. This way only parts of the original
message would be overwritten, and the original may still be
reconstructed. Of course, there is nothing that can be done if the
attacker insists on overwriting with a message size that is the
maximum embeddable in the given application. However, the computation
required to overwrite each application on a large scale might be
prohibitive enough to discourage this as a routine behaviour, at an
ISP for example.
Detection: Binaries produced by hydan should not exhibit obvious
patterns. At the most superficial level, this is accomplished by not
embedding any marker or other easily recognizable token. At best, the
embedded data looks random (which is why it is bf encrypted). At the
assembly level however, the current version of hydan makes no attempt
at mimicing the original distribution of instructions in the
application, and is thus vulnerable to statistical analysis. Indeed,
although all the instructions are equivalent, some may appear more
frequent
> steganography: the hiding of a secret message within an ordinary message and the extraction of it at its destination.
I thought steganography meant pictures of stegasaurs making little stegasarus.
Sheesh, evil *and* a jerk. -- Jade
But then I started thinking about how effectively viruses are distributed by non-techies who do click on the attachments in their EMAILs. Perhaps viruses or spyware could be used to "broadcast" a message this way to different cells in a covert organization (terrorists, organized crime, chess club members, whatever). All you'd need is an unprotected PC to act as a tethered goat and catch all those infections for later reading.
For that matter, a sender could "neuter" a virus by disabling its reproductive code and then embed a message in it and send it through some anonymizer (either a formal anonymizer or using a shell account). When the recipient stores it in a quarantine directory, it would look just like an infected EMAIL that had been cleaned up by your antivirus program, not a covert message. Some variation of this using spyware infection would be even more effective as they tend (in my limited experience) to have even more variants than viruses - the obfuscated message would be more readilly confused with normal variation. Instead of posting your tampered executables to some usenet forum, you would simply have the reciepient visit a site running the spyware. New messages would be sent and old ones sterilized when the spyware reinstalls itself.
Just my 20 mills.
"Prepare for the worst - hope for the best."
This guy wrote his assembler to generate unusual form of MOV instructions at least 10 years ago. In this way, he can find out if a program is generated using an unregistered version of A86.
Any CPU that has an instruction to exchange two registers will have some redundancy, but for X86 even basic mov (as well as add, sub, cmp and so on) specifies both two operands and a flag that specifies which one is source and which one is destination. The significance is that both operands can be registers, but only one can be a memory reference.
A much more impressive use would be a program that reads its own code as data to save the last few bytes, especially if it has a real purpose, like fitting a game into a fixed-size ROM.
I believe the program you speak of is a86. It was $50 to register it, which is why it was unpopular.
It used the same method hydan uses. It used equivelant instructions that were "different" from the way the code was written. I'd used it a bit for myself, and noticed that was what it did when I opened the files later with debug.
The documentation never really said how, it just said it "fingerprinted" the code.
The previous has been a secret message to my comrades.
Note that as far as I remember, stenography by definition is supposed to make it impossible to prove that there is data hidden there - one step further than normal encryption. It's not so much as about hiding the data as being able to deny its existance.
One reason for this is if you have encrypted data on your disk, then courts can demand the password for it. Stenography allows you to insist there is no hidden data.
A new virus is quickly spreading across the internet. Experts say it started at Defcon with a demonstration of a program that allows users to add a secret text to an executable file without altering it's filesize. Apparently the program also attached a message of it's own... don't run programs demonstrated at defcon!
I can count to 1023 on my hands. Ask me about #132.
Now, that "block all executables" setting that I can't turn find or off in Outlook will prevent terrorists from exchanging secret messages embedded in trojan executables that are attached to emails purporting to be great pornography!
It's not an annoyance; it's a *feature*!
This is a fascinating approach. One thing I didn't see mentioned at all in the documents is the possible change in performance characteristics by changing instructions which have the same effect but which have different pipeline, execution unit, or cache properties.
Modern optimizing compilers spend an awful lot of effort generating efficient combinations of instructions which try to make the most out of CPUs having complicated rules. For example, add eax,eax and shl eax,1 might both produce the same desired effect but yield significantly different runtimes depending on the presence / absence of barrel shifters or the ability of particular instructions to pair in a given CPU.
Naturally the above would only matter if the modified code is in an inner loop, but it could happen.
- determining whether a program is a virus or not is an undecidable problem (there exists no algorithm that can solve it), and
- determining whether a program is a variation of a known bounded-length virus is NP-complete (algorithms that solve the problem would take impracticably long time).
#include "/dev/tty"In cryptography, steganography has a particular meaning. In the same way that the goal of encryption is to prevent the message from being read, the goal of steganography is to prevent the message from being detected. A successful steganographic embedding is one in which a third party would not be able to find out if it is there. If you gave him two files, one with an embedded message and the other unprocessed, he should not be able to tell them apart.
For a method to truly be steganography, it's not enough just to embed some data into another. That's possible any time there's redundancy. The requirement is to make it so clever and/or subtle that there is no way to distinguish a processed file from an unprocessed one.
I doubt that this new method passes the test. Generally, while there are many synonyms possible in code, both in single instructions and in short sequences of instructions, the statistics of how these are distributed in unprocessed files are probably not random. Chances are that one synonym is used more than another. If you embed random data in a straightforward way, you will then have equal usages of both alternatives. This is a highly unusual condition, and to someone in the know, files like these will be easily distinguished.
Only if they have found a kind of synonym which already has purely random statistics, or where they are careful to precisely mimic the statistics of the original file as they add their data, can this truly be considered a form of steganography.
The documentation for the shareware DOS assembler, A86, claimed that the set of opcodes it chose to emit for various instructions was unique (i.e. those exact choices weren't made by any other assembler). Therefore if you released software assembled with A86 without registering it, if the A86 author ever got hold of that software, he'd know you had used his assembler to produce it. So the steganography in this case encoded a one bit value: "I used A86".
Pretend that something especially witty is here. Thanks.
The answer there is "you can't". You need a compiled executable large enough to have multiple instances of "alterable sequences". The way I understand it, they fiddle with reversable/interchangeable opcodes to create "bits". Say a program has 500 mixed instances of: (this is all made-up assembly)
and As you can see, a sequence of a JNZ followed by a JMP can easily be re-written as a JZ followed by a JMP. The program only needs to go through and change each instance to match bitwise value of the "message", treating JNZ-JMP as a bitwise 0 and JZ-JMP as a 1. There are probably more instances of "two ways to do it" one can exploit in a given executable to yield even bigger "message spaces".If a job's not worth doing, it's not worth doing right.
One of the most obvious examples of a jump -> jump is in the interrupt vector table, back in the DOS days. Also, some processors (such as intel 386, which I learned assembly programming on) have short and long jump instructions, so you might need something like a conditional short jump to a location that contains an unconditional long jump. IIRC, i386 didn't have conditional long jump instructions.
Did you mount a military-grade, variable-focus MASER on an unlicensed artificial intelligence?